Quantifying camouflage: how to predict detectability from ... · an object. Here, high-contrast markings intersecting an animal’s outline may be used to ‘disrupt’ the viewer’s

RESEARCH ARTICLE Open Access

Quantifying camouflage: how to predictdetectability from appearanceJolyon Troscianko1* , John Skelhorn2 and Martin Stevens1

Background: Quantifying the conspicuousness of objects against particular backgrounds is key to understandingthe evolution and adaptive value of animal coloration, and in designing effective camouflage. Quantifyingdetectability can reveal how colour patterns affect survival, how animals’ appearances influence habitat preferences,and how receiver visual systems work. Advances in calibrated digital imaging are enabling the capture of objectivevisual information, but it remains unclear which methods are best for measuring detectability. Numerousdescriptions and models of appearance have been used to infer the detectability of animals, but these models arerarely empirically validated or directly compared to one another. We compared the performance of human‘predators’ to a bank of contemporary methods for quantifying the appearance of camouflaged prey. Backgroundmatching was assessed using several established methods, including sophisticated feature-based pattern analysis,granularity approaches and a range of luminance and contrast difference measures. Disruptive coloration is afurther camouflage strategy where high contrast patterns disrupt they prey’s tell-tale outline, making it moredifficult to detect. Disruptive camouflage has been studied intensely over the past decade, yet defining andmeasuring it have proven far more problematic. We assessed how well existing disruptive coloration measurespredicted capture times. Additionally, we developed a new method for measuring edge disruption based on anunderstanding of sensory processing and the way in which false edges are thought to interfere with animaloutlines.

Results: Our novel measure of disruptive coloration was the best predictor of capture times overall, highlightingthe importance of false edges in concealment over and above pattern or luminance matching.

Conclusions: The efficacy of our new method for measuring disruptive camouflage together with its biologicalplausibility and computational efficiency represents a substantial advance in our understanding of themeasurement, mechanism and definition of disruptive camouflage. Our study also provides the first test of theefficacy of many established methods for quantifying how conspicuous animals are against particular backgrounds.The validation of these methods opens up new lines of investigation surrounding the form and function ofdifferent types of camouflage, and may apply more broadly to the evolution of any visual signal.

Keywords: Animal coloration, Background matching, Camouflage, Crypsis, Disruptive coloration, Image processing,Pattern analysis, Predation, Signalling, Vision

BackgroundAnimal coloration often plays a major role in survival andreproduction. Consequently, various forms of colora-tion—from camouflage to mimicry—have long been usedas key examples of evolution by natural selection, andhave provided an important test-bed for evolutionarythinking [1–3]. More recently, studies of colour patterns

have made substantial progress in understanding the typesof defensive coloration that exist, and the mechanismsthat make them effective [4–6]. Progress has perhaps beenmost marked in the study of concealment, with camou-flage one of the most common anti-predator strategies innature [7, 8]. However, predicting an animal’s detectabilityfrom its visual appearance remains challenging. This is animportant problem because quantifying how well ananimal avoids detection against a particular background iskey to investigating a wide range of evolutionary andecological hypotheses surrounding animal signalling and

* Correspondence: [email protected] of Exeter, Centre for Ecology and Conservation, College of Life &Environmental Sciences, Penryn Campus, Penryn, Cornwall TR10 9FE, UKFull list of author information is available at the end of the article

© The Author(s). 2017 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, andreproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link tothe Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 DOI 10.1186/s12862-016-0854-2

survival. Moreover, it will also allow us to design effectivecamouflage patterns for human use: to disguise items asdiverse as telephone masts, cameras, buildings andmilitary equipment and personnel. Indeed, rigorouslyquantifying colour patterns in general has been a topic ofconsiderable recent interest due to both the potentialhuman applications and the unique opportunity to revolu-tionise our understanding of animal signalling [9–11].Regarding camouflage, identifying and assessing the effect-iveness of a wild animal’s camouflage strategy is essentialfor understanding predator–prey interactions in anysystem with visually guided predators. Many conspicuousdisplays will also be influenced by evolutionary pressurefor greater camouflage, for example, displays that areaposematic at close quarters can be cryptic from furtheraway [12], and conversely, an understanding of howcamouflage is achieved can illuminate the mechanismsthat make conspicuous coloration effective [13]. Quantify-ing an animal’s appearance relative to its background istherefore essential for investigating sexual, aposematic andcamouflaged displays in a diverse range of fields.A number of different types of camouflage have been

identified, based on how they hinder detection. Themost ubiquitous camouflage strategy is probablybackground matching, where animals match the generalcolour, brightness and patterns of their backgrounds [8,14]. However, one of the key features thought to facili-tate detection and/or recognition is the overall outline ofan object. Here, high-contrast markings intersecting ananimal’s outline may be used to ‘disrupt’ the viewer’sability to discern or detect it [4, 7, 15], a strategy calleddisruptive coloration. The most direct way to determinean animal’s camouflage, and how effective it is, usesoften lengthy behavioural tests or survival experimentsthat are difficult to undertake in the wild [16]. Consider-able effort has therefore pursued computer models thatcan quantify how well any two visual samples matchaccording to different processing or camouflage theories[17–22]. However, these camouflage models have rarely,if ever, been directly compared under controlledconditions, nor using data on observer success in findinghidden objects. This lack of model validation means thatresearchers rarely know which methods they shouldadopt when investigating an animal’s appearance.Furthermore, models that quantify visual signals andtheir match (or contrast) with the background have thepotential to greatly inform us regarding the mechanismsthrough which colour patterns work, and how theyshould be optimised for maximal success (or indeed,traded-off with other competing functions). If modelsthat are based on specific or generalised features ofvisual processing fit with behavioural data, this can illu-minate the possible mechanisms through which colourpatterns are made effective [19], and even how changes

to them might improve the adaptive value of the de-fence. Where the models are inspired by known or likelyneural architecture this can even reveal likely avenuesfor research into the underlying structures performingthe visual processing.Here we set out to address the above issues by

pitting numerous contemporary models of camouflagedirectly against one another, using human observers tocompare models to detection times. We tested sevenclasses of model that have been previously used forinvestigating background matching and disruptivecoloration hypotheses. These models compare prey totheir backgrounds according to three different criteria: i)luminance matching, ii) pattern matching and iii) disrup-tive coloration. Luminance (i.e. perceived lightness)matching is the simplest form of camouflage to measure,for example calculating the average difference in lumi-nance between prey and their backgrounds, which isthought to be important in explaining survival inpeppered moths against mismatching backgrounds [23].Contrast differences measure how similar the variation inluminance in the prey is to that of the prey’s background.This has been found to be important in the survival ofwild ground-nesting bird clutches [16]. Pattern matchingis another important aspect of background-matchingcamouflage that has recently been found to predict thesurvival of nightjar nests [16]. Pattern matching has beenmeasured using a number of different methods that varyin their biological plausibility and complexity. Thesemethods generally separate images into a series of spatialscales (e.g. using fast Fourier transforms, or Gabor filters),then compare the information at these different scalesbetween the prey and background. Other methods searchfor shapes or features found in the prey that are similar tothose in their backgrounds [22, 24]. For an overview seeTable 1 and Additional file 1. The final type of camouflagemeasured was disruptive coloration, where contrastingmarkings break up an object’s outline and create falseedges [7, 8, 15]. This camouflage strategy has receivedconsiderable investigation in the last decade, and has beenshown to be highly successful in numerous contexts, in-cluding where stimuli have various contrast and patternfeatures manipulated [4, 25–30]. In spite of the clear pro-tective benefits of artificial disruptive edges, it has provenfar more difficult to measure how disruptive real prey areagainst their backgrounds [19, 31]. Recent edge disruptionmeasures have quantified how many edges are present inthe animal’s outline [32], or have used the Canny edgedetector to measure the number of perceived edges in theprey’s outline relative to its surroundings [21], or the num-ber of outline edges relative to those in the object’s centre[33]. However, these measures do not take into accountthe direction of the perceived edges, so cannot distinguish‘false edges’ (that run at right angles to the prey’s outline

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 2 of 13

and should be maximally disruptive [8, 31]) from ‘coherentedges’ that match the angle of the animal’s outline, makingthe prey’s tell-tale shape easier to detect [34]. We thereforedeveloped a novel edge disruption metric called ‘GabRat’that uses biologically inspired and angle-sensitive Gaborfilters to measure the ratio of false edges to coherent edgesaround a target’s outline (see Fig. 1e) [35–37]. A high ratioof false edges to coherent edges should be more disrup-tive, making prey more difficult to detect. Backgroundcomplexity is also known to influence detection times[38]. For example, artificial moths were more difficult tofind for humans and birds alike when the surroundingbark had higher levels of luminance contrast and edgeorientation changes [24]. However, in this study we aimedto focus on metrics that investigate the interaction be-tween target and background, rather than assess the gen-eral properties of backgrounds that affect concealment.We tested how the above camouflage models predicted

the performance of human ‘predators’ searching forcamouflaged stimuli against natural background imageson a touch screen monitor. Each prey was unique, gener-ated from its background using methods that maximisedor minimised the prey’s level of edge disruption, with preyalso varying in their level of background matching. Weused tree-bark backgrounds as these are biologicallyrelevant backgrounds for a wide range of camouflagedanimals, and they exhibit a degree of background

heterogeneity in contrast and spatial features. Artificialprey and tree-bark backgrounds such as these have beenused extensively for testing camouflage theories becausethey capture the essence of camouflage patterns effectivelywithout the need to find and take calibrated images oflarge numbers of camouflaged prey [4, 27, 28, 32]. Thesestudies have also demonstrated that human and non-human visual systems respond similarly to thesecamouflage stimuli. We calculated the preys’ camouflagewith a battery of different models to determine which bestpredicted capture times. Each prey’s level of camouflagewas measured between their entire background image, orwith the local region within one body length to investigatewhether camouflage is best predicted by local or globalbackground matching. In addition we tested forinteractions between the most effective luminance-, pat-tern- and disruption-based camouflage metrics to deter-mine whether extreme luminance differences renderpattern and disruption strategies ineffective. Finally, wediscuss the effectiveness of the most commonly usedmodels for assessing an animal’s camouflage and what ourfindings reveal about the mechanisms underlying animalcamouflage.

ResultsThere were substantial differences between the abilitiesof different camouflage metrics to predict capture times,

Table 1 Descriptions of the methods used to measure prey conspicuousness

CamouflageCategory

Variable Name Filtering Method Basic Description

Edge Disruption GabRat Gabor Filter Average ratio of ‘false edges’ (edges at right angles to the preyoutline) to ‘salient edges’ (edges parallel with the prey outline).See Additional file 1

VisRat Canny Edge Detector Proportion of Canny edge pixels in the prey’s outline region [21]

DisRat Canny Edge Detector Proportion of Canny edge pixels in the prey’s outline region [34]

Mean Edge-regionCanny Edges

Canny Edge Detector Proportion of Canny edge pixels in the prey’s outline region.

Edge-intersectingcluster count

None Count of the number of changes in the pattern around the prey’soutline [33]

Pattern/Objectdetection

SIFT Difference-of-Gaussians,Hough Transform

Uses Hough transform to find features in the prey, then countshow many similar features are found in the background [19, 22]

HMAX Gabor Filter Breaks down a bank of Gabor Filter outputs into layers thatdescribe patterns with some invariance to scale and orientation [20]

Pattern PatternDiff Fourier Bandpass Sums the absolute difference between the prey’s pattern statistics [42]

Euclidean PatternDistance

Fourier Bandpass Euclidean distance between normalised descriptive pattern statistics[42]

Luminance Mean BackgorundLuminance

Luminance Mean luminance

Mean LuminanceDifference

Luminance Absolute difference between mean prey and mean backgroundluminance

LuminanceDiff Luminance Sum of absolute luminance histogram bins [16]

Contrast Difference Luminance Absolute difference of contrast between prey and background,where contrast is the standard deviation in luminance levels

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 3 of 13

see Fig. 2 for full results. Camouflage experiments suchas this are expected to entail a very high levels of re-sidual variation in capture times due to the interactionbetween the location of the prey, the viewers’ eye

movements [32], and the heterogeneous nature of thebackgrounds. For example, prey that appeared in thecentre of the screen were captured sooner than thosenearer the edges, explaining 8.75% of model deviance in

Fig. 1 Examples of prey and edge disruption measurements. (a) Sample prey highlighted in blue against its background image. The ‘local’ regionwithin a radius of one body-length is highlighted in red. (b) Examples of prey generated with the disruptive algorithm (left) and background-matchingalgorithm (right). These prey were chosen as their GabRat values were near the upper and lower end of the distribution (see below). (c) Illustration ofthe GabRat measurement. Red and yellow false colours indicate the perceived edges run orthogonal to the prey’s outline (making disruptive ‘falseedges’), blue false colours indicate the perceived edges are parallel to the prey’s outline (making ‘coherent edges’). GabRat values are only measuredon outline-pixels, so these values have been smoothed with a Gaussian filter (σ = 3) to illustrate the approximate field of influence. The prey on the lefthas a high GabRat value of 0.40, while the prey on the right has a low GabRat value (0.20). (d) Canny edges are highlighted in the images. Edges insidethe prey are highlighted in blue, edges in the prey’s outline region are green, and edges outside the prey are red. The VisRat and DisRat disruptionmetrics are formed from the ratios of these edges. (e) Gabor filter kernels (sigma = 3), shown in false colour at the different angles measured

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 4 of 13

the best model, while the random effects of participantID and background image ID explained 3.16 and 1.68%of variance respectively. The best predictor of capturetimes was GabRatσ3, which measured edge disruptionand explained 8.34% of model deviance (p < 0.001). Asan illustration of effect size, prey in the upper 10% ofGabRatσ3 values took on average 2.23 times longer tocatch than those in the lower 10% of GabRatσ3 values(4.447 s as opposed to 1.998 s). This was followed by theLocal SIFT model (measuring pattern/feature similar-ities, explaining 5.65% deviance, p < 0.001), and togetherwith other GabRat sigma values (which specify the sizeof the filter), these were the only metrics that performedbetter than the prey treatment (i.e. whether it was gener-ated using the background matching or disruptivealgorithm, which explained 3.71% deviance, p < 0.001).The worst metric at predicting capture times was theCanny edge disruption measurement DisRat, explainingless than 0.01% of model deviance (p = 0.879), althoughthis was likely due to its non-linear nature, see below.The full model comparing the best edge disruption,

pattern and luminance predictors contained GabRatσ3,Local SIFT difference and Mean Local LuminanceDifference metrics. Following AIC simplification themodel retained an interaction between GabRatσ3 andSIFT local that explained 0.21% deviance, with the maineffect of GabRatσ3 explaining the majority of deviance(8.18%) and SIFT local with (3.05%) all terms were sig-nificant (p < 0.001). The global comparisons model based

on bandpass descriptive statistics performed compara-tively well, explaining 1.87% of deviance when summedacross all model terms. This model contained four two-way interactions that retained all five descriptivevariables (full model output is available in Additionalfile 1). The local comparisons model using bandpassdescriptive statistics performed less well, retaining justDominant Spatial Frequency Difference as a predictorthat explained 0.42% of deviance). While backgroundcomplexity measured independently of the prey wasnot the focus of this study, a number of metricseffectively include this information, such as the Meanedge-region Canny edges (deviance = 0.21%, p = 0.001),and Mean Local Bandpass Energy (deviance = 0.19%,p = 0.002).Gabor-based pattern-matching metrics did not vary

consistently between local and global difference predic-tors. The bandpass-based pattern matching metricsperformed better when comparing the prey to their globalregion than their local region with the exception of theEuclidean Pattern Distance, which performed betterlocally. In contrast, luminance metrics all performed bet-ter when considering the local rather than global regions.However this is perhaps to be expected given the way thebackground images were normalised, and the way preywere generated from their backgrounds. Nevertheless, theGlobal PatternDiff metric performed substantially betterthan the Global Mean Luminance Difference, which aspredicted is non-significant (deviance = 0.04%, p = 0.143).

Fig. 2 Capture time prediction accuracy. The predictive performance of camouflage metrics tested in this study ranked from best to worst. Allcamouflage metrics were continuous variables using one degree of freedom in each model with the exception of treatment type, which wascategorical, consuming two degrees of freedom. Note that DisRat and VisRat performed better when fitted with a polynomial

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 5 of 13

Given the striking difference in performance of DisRatand GabRat metrics we tested how well each of them pre-dicted prey treatment. As predicted, disruptive prey had asignificantly lower DisRat and higher GabRatσ3 thanbackground-matching prey (linear model; DisRat: F1, 3817= 1413, p < 0.001; GabRatσ3: F1, 3817 = 708.2, p < 0.001),demonstrating that both were able to predict treatmenttype. When fitted with a quadratic, VisRat local andDisRat both fitted capture times significantly better (basedon model comparison, p < 0.005), increasing the devianceexplained by these variables to 0.439 and 0.558% respect-ively. The optimal VisRat local ratio was equal to 0.951,while the optimum DisRat was 0.903, values higher orlower resulted in shorter detection times.

DiscussionThe number of studies quantifying the appearance ofanimals to test evolutionary and ecological hypotheses isincreasing rapidly with the advancement of imagingmethods, computer models and models of animal vision[9–11]. However, the methods developed to determinehow conspicuous an animal is against its backgroundhave rarely been validated using behavioural data, letalone compared to alternative models. This is an issuethat goes beyond simply developing the best techniquesto quantify animal appearances; coupling visual modelsto performance, and determining which metrics aremost effective regarding observer behaviour, can alsoenable predictions about the optimisation of visualsignals in nature and in applied tasks. By comparing theperformance of a suite of different analysis techniqueswe have determined the best methods for quantifyingdetectability from appearance.We found that there were striking differences between

the abilities of different camouflage metrics to predictthe capture times of our computer-generated prey. Ourstudy broadly validates the use of the majority of camou-flage models used in the literature to date, however therewere important differences and exceptions, demonstrat-ing the importance of behavioural validation of thesemodels. The Gabor Edge Disruption Ratio (GabRat) de-vised for this study performed substantially better thanall other metrics; prey with high GabRat values were halfas likely to be captured by our human ‘predators’ in agiven time-frame than those with low values, demon-strating the potential for powerful evolutionary selectionpressures on this metric. Moreover, GabRat was overtwice as effective at predicting capture times as the typeof algorithm used to generate the prey, and over tentimes better than Fourier bandpass, HMAX or lumi-nance difference metrics. Also striking was the relativefailure of Canny edge detection-based models (e.g. Vis-Rat and DisRat) to predict capture times when treated aslinear predictors (i.e. testing the hypothesis that lower

VisRat or DisRat values result in longer capture times).When VisRat and DisRat were fitted non-linearly, theoptimal ratios were slightly below one in both cases,where ratios equal to one would fit with a background-matching strategy, and ratios below one are disruptive.The non-linear performance of VisRat and DisRat makethem much more difficult to use as predictors of detect-ability without considerable further investigation of theoptimal ratio, which may even change between studysystems. The fact that the optimal VisRat and DisRatvalues were close to one suggests that they are eithernot measuring edge disruption effectively, or that theoptimal level of disruption is very close to abackground-matching strategy (which is contradicted bythe GabRat result). DisRat was, however, a goodpredictor of treatment type, able to determine whetherprey were generated using the background matching ordisruptive algorithms slightly better than GabRat. Thisdemonstrates that the Canny edge methods were notfailing due to misidentification of edge artefacts on theprey’s outline. In line with our predictions based onbiases in human spatial frequency detection [39], GabRatwas most effective with a sigma of 3 pixels. This alsosuggests that the Canny edge metrics should have beenperforming optimally for the viewing distance, as theywere also calculated at this scale. Taken together thissuggests that the angle of the perceived edges relative tothe prey’s outline is essential in any model attempting todescribe edge disruption, as this is the primary differencebetween the Canny and GabRat methods that werefound to behave so differently in this study. The Canny-edges shown in Fig. 1d demonstrate why basing metricson the presence of edges alone is not sufficient; the dis-ruptive prey in this example has a large number ofdetected edges in its outline region that mostly run atright angles to the outline.The success of GabRat in predicting capture times is

all the more striking given the comparatively small areathat it measures (e.g. see Fig. 1c). The local comparisonzone encompassed approximately 92,000 pixels, and theglobal camouflage metrics measured 2.1 million pixels.By contrast, the GabRatσ3 kernel has a maximum diam-eter of just 19 pixels, covering an area of approximately5500 pixels. Even though GabRat only takes into account0.26% of the available data in the scene, those data werefound to be far more effective for predicting capturetimes than any other data we measured, supporting thenotions of Cott [7] and Thayer [15] that the animal’soutline tends to give it away, and suggesting our workingdefinition of edge disruption that takes into account thedifference between perceived edges and body outline [8]fits with the observed data. In addition, GabRat is one ofthe least computationally demanding metrics measuredhere, and uses biologically inspired methods. The

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 6 of 13

variables required for specifying the model can be basedon neurophysiology [20] without the need for guessingvariables and thresholds, which sets it apart from theCanny or SIFT-based models [18], or edge-intersectingpatch counts [32]. An alternative conclusion is that thepattern and luminance-based metrics we have measuredare less effective, perhaps because they fail to model realvisual systems adequately, although these methods havea substantial track record in supporting hypotheses innatural systems [16, 22, 40, 41].In line with Webster et al. [32], we found the Edge-

Intersecting Patch Count was a good predictor ofcapture time, indeed it outperformed all pattern- andluminance-based metrics other than SIFT even though itis blind to the interaction between prey and their back-grounds. However, it is also a less useful metric for gen-eralising to other systems where the edge-intersectingpatches are less easily defined. For example, how shoulddiscrete patches be distinguished in real prey images,what regions around the prey’s outline should be used,and what visual processing architecture could reproducethese counts? Therefore, although this metric is success-ful in this study where patches are defined by the preygeneration algorithm, we think it an unlikely avenue forfruitful future research into edge disruption compared tometrics that more closely match known neural process-ing methods.Contrary to our expectations based on Moreno et al.

[42], the best method of quantifying pattern differencewas the SIFT model. Although in line with our predic-tions, prey took longer to capture if they shared morefeatures with their local background region than theirglobal background. This result is similar to experimentsdemonstrating that captive blue tits Cyanistes caeruleustook longer to find prey against backgrounds with higherdensities of geometric shapes identical to those on theprey [43]. Our finding suggests that the same effectholds true when natural backgrounds rather than re-peated geometric shapes are used. The SIFT model wasalso the only pattern matching model that performedbetter than treatment type, which is perhaps surprisinggiven treatment type was blind to the interaction be-tween individual prey and their backgrounds. As pre-dicted, the HMAX models performed better than theFourier-based bandpass models. The HMAX modelsthat forced comparisons to be made between prey andbackground without allowing for orientation changeswere more effective, for example demonstrating thatwhere there were stripes on the prey these offered themost effective camouflage when they were at the sameorientation as the stripes in their background at a similarspatial scale. The Fourier-based global PatternDiff metricperformed comparatively well compared to the HMAXmetrics even though it is substantially less computationally

demanding and less biologically accurate. The otherFourier-based metrics fared less well, although when theglobal pattern descriptive statistics were combined into anoptimal model it predicted capture times well, indeedperforming better than any other HMAX or Fourier-basedmethod. However, this model is not directly comparable tothe others in that it was adapted to fit the observed data inthis study from a large number of degrees of freedom,giving it an unfair advantage. Nevertheless, this process isuseful because it highlights those descriptive pattern met-rics that best predicted capture times in this study, makingit the most useful method for producing information onthe importance of different aspects of pattern matching,such as whether spatial scale or energy measures are moreimportant. By contrast the SIFT model provides the leastinformation on what aspects of the general patterns aremost important, making it less easy to infer what types offeatures are most important, how their importance isweighted, and whether these variables and weightings applyequally well to non-human visual systems.Our data suggest that while matching the average

background luminance is important, it is substantiallyless important than pattern matching or edge disruptionmetrics. We might have expected to find that patternand edge disruption should only be important up thepoint where the prey become so different in average lu-minance to their background that they stand out(switching from inefficient to efficient search strategies[34]). However, the best luminance predictor (LocalMean Luminance Difference) was dropped from the finalmodel of the best predictors, suggesting that this is notthe case. Nor was there autocorrelation between thisluminance metric and the best pattern and edge disrup-tion metrics, demonstrating—contrary to our expecta-tion—that prey can mismatch the luminance of theirlocal background and still have a good SIFT patternmatch and level of GabRatσ3 edge disruption. Prey inreal-world situations could have a level of luminancemismatch with their environment beyond those achievedby our computer display, however most backgroundmatching prey would not be expected to have such a bigluminance difference to their background. The inter-action in the final model of best predictors betweenGabRatσ3 and Local SIFT pattern match suggest thesemetrics can operate in synergy to increase detectiontimes. Although, the effect size of this interaction wassmall compared to the abilities of GabRatσ3 and SIFT topredict capture times on their own.To our knowledge this study is the first to compare a

wide range of different contemporary methods for test-ing levels of achromatic camouflage. We have validatedthe use of GabRat, a novel edge disruption metric, whilethe VisRat and DisRat metrics adopted in the literatureto date for investigating edge disruption cannot be used

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 7 of 13

as reliable indicators of detectability. Fourier-basedmethods were less effective than more complex and bio-logically plausible methods, they were, however, themost informative for distinguishing different aspects of apattern’s nature, and were still highly significant predic-tors of capture time. We would still therefore recom-mend their use in cases where little is known about thereceiver’s visual system. HMAX models, while being themost biologically plausible for quantifying pattern differ-ence were not found to be as effective as SIFT forpredicting capture times, indicating that the number ofdistinct features shared between two patterns is moreimportant than the overall scales, angles and contrastlevels. Our use of tree-bark backgrounds photographedunder diffuse lighting conditions may also have influ-enced our findings, and qualitatively different resultscould be possible against alternative backgrounds, andunder different lighting conditions. A number of studieshave demonstrated the importance of background com-plexity in affecting detectability [24, 38, 43], so our find-ings may not hold against simple backgrounds with lowlevels of achromatic complexity. Xiao and Cuthill foundthat feature congestion best predicted human and birddetection of artificial moths [24]. This metric combineslocal achromatic and chromatic changes, and edgeorientation changes. While our study did not considercolour differences or feature congestion explicitly, itmeasured a number of variables similar to those used incalculating the achromatic components of feature con-gestion; for example by measuring the number of Cannyedges locally, analysing the local pattern energy, andHMAX Gabor filtering, which takes into account edgeorientations. While we found that all of these metricspredicted capture times in line with Xiao & Cuthill, theywere not as effective as other methods, possibly becausethey do not consider the prey’s interaction with its back-ground. Future work should compare the effectivenessof these models with images of natural prey, and inwholly natural systems to establish how wide-rangingthese findings are to detection times in alternativecontexts and with non-human visual systems [24]. Inaddition, models should be developed that can integratechromatic cues; experiments on colour discriminationtypically involve comparisons between flat colourpatches rather than the complex and varied colours en-countered in natural search tasks.

ConclusionsThis study demonstrates how best to measure camou-flage from appearance, however the same models canalso be used to measure signals that stand out from thebackground [13]. The methods tested in this study aretherefore useful for researchers studying the appearanceof wide-ranging signals, from sexual and aposematic

displays to mimicry and camouflage in fields from evolu-tionary and sensory ecology to military camouflage andadvertising. Model validation in humans can also help toreduce the number of costly animal behaviour experi-ments required for testing visual hypotheses. Ourfindings have two main evolutionary implications: first,that we would expect camouflage to be optimised bycreating false edges at scales linked to the typicaldetection distances of the receiver, and second, thatwhile visual systems should have evolved to overcomethis weak-spot as effectively as possible, recognising theanimal’s outline is still key to detection and/or recogni-tion. Likewise, signals that have evolved to maximise de-tection (such as sexual or aposematic displays, [13])should do the opposite, creating coherent edge cues atscales relevant to typical viewing distances.

MethodsThe performance of different models of camouflagemeasurement was assessed in human touch-screenpredation experiments. Although animal vision variessubstantially between taxa, human performance intouch-screen experiments has been found to agree withbehavioural data from non-humans [27]. Furthermore,spatial visual processing is thought to be similar betweentaxa [17], suggesting the results based on achromatic hu-man performance should be good indicators of perform-ance in many other vertebrate species.

Backgrounds and prey generationPhotographs of natural tree bark were used as back-ground images (oak, beech, birch, holly and ash, n = 57images), taken using a Canon 5D MKII with a NikkorEL 80 mm lens at F/22 to ensure a deep depth of field.Photographs were taken under diffuse lighting condi-tions. Luminance-based vision in humans is thought tocombine both longwave and mediumwave channels [44].As such we used natural images that measure luminanceover a similar range of wavelengths by combining thecamera’s linear red and green channels. Next the imageswere standardised to ensure that they had a similar over-all mean luminance and contrast (variance in luminance,see [16]). Images were then cropped and scaled with a1:1 aspect ratio to the monitor’s resolution of 1920 by1080 pixels using bilinear interpolation. Images werelog-transformed, resulting in a roughly normal distribu-tion of luminance values. A histogram of logged pixelvalues with 10,000 bins was analysed for each image.The 1st, 50th (median) and 99th percentile luminancevalues were calculated, and their bins modelled with aquadratic function against the desired values for thesepercentiles to ensure the median was half way betweenthe luminance at the upper and lower limits. The result-ing images all have an approximately equal mean and

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 8 of 13

median luminance, and similar luminance distributions(contrast), and equal numbers of pixels at the upper andlower extremes.Each prey was generated from the background it was

presented against using custom written code similar tothat used previously [28]. This methodology createsunique two-tone prey that match the general patternand luminance of the background (see Fig. 1b). Briefly,for each prey a triangular section of the backgroundimage was selected from a random location, 126 pixelswide by 64 pixels high. For disruptive prey the thresholdlevel was calculated that would split the image into thedesired proportions (60% background to 40% pattern).For background matching prey a Gaussian gradient wasapplied to the edges prior to threshold modelling thatmade it less likely that underlying patterns would comethrough nearer the edge of the prey. This avoids creatingsalient internal lines in the background matching preyparallel with the prey’s outline, while ensuring no pat-terns touch the very edges. If the thresholded proportionwas not within 1% of the target limits the process wasrepeated. Prey were generated with either dark-on-lightor light-on-dark patterns, and each participant onlyreceived one of these treatments. Dark-on-light prey hadthe dark value set to equal the 20th percentile of thebackground levels and the light value set to the 70thpercentile. The light-on-dark prey used the 30th and80th percentiles respectively. The differences betweenthese two treatments are due to the fact that there isslightly more background area than pattern, and thesevalues ensure that the overall perceived luminance ofthe two treatments is similar to the median backgroundluminance, factoring in the 60/40 split of background topattern area.

Calculating camouflage metricsThe camouflage metrics measured in this study fall intoseven distinct methodologies, though many of these inturn provide a number of additional variations: Gaboredge disruption ratios (GabRat, first proposed in thisstudy), visual cortex-inspired models based on theHMAX model [20, 45], SIFT feature detection [18, 22],edge-intersecting patch count [32], luminance-basedmetrics [10, 16], Fourier transform (bandpass) patternanalysis [10, 16, 41], and edge-detection methods toquantify disruption [19, 21, 33]. Where possible, we haveused the same terminology for the different metrics asthey are used in the literature. Many of these variablescan be used to compare a prey target to a specific back-ground region. Therefore, where the metrics allowed, wecompared each prey to its entire background image (the‘global’ region) and to its ‘local’ region, defined as thearea of background within a radius of one body-length(126 pixels) of the prey’s outline. The distance of one

body length is the largest ‘local’ area that would not ex-ceed the background image limits, because prey were al-ways presented within one body length of the screenedge. This also ensured that the shape of the local regionwas always uniform, however one body length is also aflexible unit of scale measurement that could be used inother animal systems. Measuring two regions allowed usto test whether a prey’s local or global camouflagematching was more important across the different met-rics (see Fig. 1a). If the prey are a very poor luminancematch to their backgrounds then we might expect themto stand out enough for the comparatively low acuityperipheral vision to detect them easily using efficientsearch [34]. However, where the prey are a good lumi-nance and pattern match to their local background theyshould be most difficult to detect as this would requirethe participant adopt inefficient search strategies, slowlyscanning for the prey. We can make further predictionson the importance of pattern and edge disruption at spe-cific spatial scales given humans are most sensitive tospatial frequencies in the region of around 3–5 cyclesper degree [39]. This scale is equivalent to a Gabor filterwith a sigma between approximately 2–4 pixels.For clarity we use the term ‘edge’ to refer to perceived

edges based on a given image analysis metric, and ‘out-line’ to refer to the boundary between prey and back-ground. Unless otherwise specified, these methods wereimplemented using custom written code in ImageJ. TheGabRat implementation will be made available as part ofour free Image Analysis Toolbox [10], code for all othermetrics is already available in the toolbox, or is availableon request. See Table 1 for an overview of the measure-ment models.

Gabor edge disruption ratio (GabRat)Prey were first converted into binary mask images (i.e.white prey against a black background), a Gabor filterwas then applied to each of the pixels around the edgeof the prey at a range of angles (four in this study; theGabor filter settings were identical to those used in theHMAX model below, and Fig. 1e). The angle of theprey’s outline at each point (parallel to the outline) wasthe angle with the highest absolute energy (|E|) mea-sured from the mask image. Each point around theprey’s outline in the original image was then measuredwith a Gabor filter at an angle parallel to, and orthog-onal (at right angles) to the edge at each point. Thismeasured the interaction between the prey and its back-ground. The disruption ratio at each point on the prey’soutline was then calculated as the absolute orthogonalenergy (|Eo|) divided by the sum of absolute orthogonaland absolute parallel energies (|Ep|). Finally, the Gaboredge disruption ratio (GabRat) was taken as the mean ofthese ratios across the whole prey’s outline:

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 9 of 13

GabRat ¼Σ Eoj j

Eoj jþ Epj jð Þn

Consequently, higher GabRat values should imply thatprey are disruptive against their backgrounds (having ahigher ratio of false edges), and lower GabRats implythat the edges of prey are detectable (see Fig. 1c). Thisprocess was repeated with different sigma values for theGabor filter to test for disruption at different spatialfrequencies (sigma values of 1, 2, 3, 4, 8 and 16 weremodelled in this study). It is therefore possible for preyto be disruptive at one spatial scale or viewing distance,while having more detectable edges at another.

HMAX modelsThe HMAX model is biologically inspired, based on anunderstanding of neural architecture [20]. It breaksdown images using banks of Gabor filters [37] that arethen condensed using simple steps into visual informa-tion for object recognition tasks. It was also found tooutperform the SIFT in an object classification compari-son [42], so we might therefore expect it to perform bestin a camouflage scenario. The HMAX model was devel-oped in an attempt to emulate complex object recogni-tion based on a quantitative understanding of the neuralarchitecture of the ventral stream of the visual cortex[20, 45]. Our HMAX model first applied a battery of Ga-bor filters to the prey image, and the local and globalbackground regions. The Gabor filters were applied atfour angles and ten different scales, with Gamma = 1,phase = 2π, frequency of sinusoidal component = 4,minimum sigma = 2, maximum sigma = 20, increasing insteps of 2. C1 layers were created following Serre et al.[20] by taking the maximum values over local position(with a radius of sigma + 4) and the neighbouring scaleband. The mean of each scale band in the prey’s C1 wasthen calculated as we wished to compare the prey’s over-all pattern match rather than a perfect template match,which would test a masquerade rather than background-matching hypothesis [6]. This C1 template was thencompared to each point in the local surround and entire(global) background image. The average match betweenthe prey’s C1 layer and the C1 layers of each region wassaved along with the value of the best match and thestandard deviation in match (a measure of heterogeneityin background pattern matching). This model was runboth with and without an allowance for the template torotate at each point of comparison. When rotation wasallowed the angle that had the best match at eachcomparison site was selected. The HMAX model withrotation describes how well the moth’s average pattern(i.e. angles, scales and intensities) matches the back-ground if the prey can rotate to the optimal angle at

each point. The model without rotation forces the preyto be compared to its background at the same angles, soshould be a better predictor in this study where theprey’s angle relative to its background is fixed.

SIFT feature detectionScale Invariant Feature Transform (SIFT, [18]) modelswere primarily developed for object recognition and rap-idly stitching together images by finding sets of sharedfeatures between them even if they occur at differentscales or angles. Although the SIFT models share somesimilarities with known biological image processing andobject recognition at certain stages, such as the inferiortemporal cortex [46], the method as a whole is notintended to be biologically inspired, although it has beenapplied to questions of animal coloration [22].The SIFT function in Fiji (version 2.0.0 [47]) was used

to extract the number of feature correspondences be-tween each prey and its local and global backgroundwithout attempting to search for an overall templatematch. Settings were selected that resulted in a largeenough number of correspondences that the count datawould exhibit a normal rather than Poisson distribution,but that was not too slow to process. These settings pro-duced roughly 300 features in the prey, and 300,000 inthe background in a sub-sample run to test the settings.The initial Gaussian blur was 1, with 8 steps per scaleoctave, a feature descriptor size of 4, 8 orientation binsand a closest to next-closest ratio of 0.96. Prey and theirlocal background regions were measured in their entir-ety against a white background. As it stands this meth-odology might not therefore be suitable for comparingprey of different shapes and sizes without further modifi-cation and testing.

Edge-intersecting cluster countThe number of cases where patterns intersected thepreys outline (following [32]) were summed in each preyusing a custom written script. Background matchingprey had no instances of edge intersections, which wouldcreate zero inflation and violate model assumptions. Wetherefore analysed a subset of the data containing onlydisruptive prey for testing this metric.

Luminance-based metricsPrey were compared to their local and global back-ground regions using a number of luminance-based met-rics that could affect capture times. Luminance wastaken to be pixel intensity values. LuminanceDiff wascalculated as described in Troscianko et al. [10], as thesum of absolute differences in the counts of pixel num-bers across 20 intensity bins, essentially the difference inimage luminance histograms. This measure is suitable

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 10 of 13

where the luminance values do not fit a normal distribu-tion, which is the case with our two-tone prey. Meanluminance difference was the absolute difference inmean luminance values between prey and backgroundregions. Contrast difference was calculated as the abso-lute difference in the standard deviation in luminancevalues between prey and background region. Mean localluminance was simply the mean pixel level of the localregion. This was not calculated for the entire back-ground image because they had been normalised to havethe same mean luminance values.

Bandpass pattern metricsFourier Transform (bandpass) approaches [17, 40] onlyloosely approximate the way visual systems split animage into a number of spatial frequencies, however,they have a long and successful track record of use inbiological systems, are fast to calculate and provide out-put that can be used flexibly to test different hypotheses[16, 41]. Fast-Fourier bandpass energy spectra were cal-culated for the prey and their local and global back-ground regions using 13 scale bands, increasing from2px in multiples of √2 to a maximum of 128px [10]. Pat-ternDiff values were calculated as the sum of absolutedifferences between energy spectra at each spatial band[10]. This metric describes how similar any two patternsare in their overall level of contrast at each spatial scale.Descriptive statistics from the pattern energy spectrawere also calculated, these being: maximum energy,dominant spatial frequency (the spatial frequency withthe maximum energy), proportion power (the maximumenergy divided by the sum across all spatial frequencies,mean energy and energy variance (the standard deviationin pattern energy, a measure of heterogeneity acrossspatial scales) [10, 41]. A metric similar to the multidi-mensional phenotypic space used by Spottiswoode andStevens [48] was calculated from these descriptive statis-tics. However, rather than sum the means of eachdescriptive pattern statistic, a Euclidean distance wascalculated after normalising the variables so that eachhad a mean of zero and standard deviation of one(ensuring equal weighting between pattern statistics).We termed this metric ‘Euclidean Pattern Distance’. Inaddition, a full linear mixed model was specifiedcontaining all descriptive pattern statistics and all two-way interactions between them (the model was specifiedin the same form as other linear mixed models in thisstudy, see statistical methods below). The full model wasthen simplified based on AIC model selection.

Canny edge detection methodsCanny edge detection methods were applied followingLovell et al. [21] and Kang et al. [33], using a Java imple-mentation of the method [49]. The Canny edge filter

was applied to each image with the settings specified byLovell et al., using a sigma of 3 and a lower threshold of0.2. The upper threshold required by the Canny edgedetection algorithm was not specified by Lovell et al., soa value of 0.5 was selected that ensured there would beno bounding of the data where no edges were detected.Following Lovell et al., the prey’s outline region wasspecified as being four pixels inside the prey’s outlineand 8 pixels outside (see Fig. 1d). As above, two back-ground regions were measured; local and global,although the 8px band around the prey’s outline was notincluded in the local or global regions. We measured themean number of Canny edge contours in each region(i.e. the number of edge contour pixels in each regiondivided by the total number of pixels in each region tocontrol for the differences in area being measured). It isunclear whether Lovell et al. applied this control, how-ever given the areas being measured are fixed in thisexperiment (all prey and backgrounds are the same size)this would not affect the overall outcome. VisRat wascalculated as the mean Canny edge contours found inthe background region (either local or global) divided bythe mean Canny edge contours found in the prey’s out-line region (termed ContEdge by Kang et al.). DisRatwas calculated following Kang et al. as being the meanCanny edge contours found inside the prey (termedMothEdge by Kang et al.) divided by ContEdge. BothVisRat and DisRat required a log transformation to dem-onstrate a normal error distribution.

Experimental setupPrey were presented at a random location against theirbackground using custom written HTML5/Javascript codeon an Acer T272HL LCD touch-screen monitor. The dis-play area was 600 mm by 338 mm, 1920 by 1080 pixels.The monitor’s maximum brightness was 136.2 lux, andminimum was 0.1 lux, measured using a Jeti Specbos 1211spectroradiometer. The monitor’s output fitted a standardGamma curve where brightness (lux) = 8.362E-4*(x +25.41)^2.127*exp (−(x + 25.41)/3.840E11), where x is an 8-bit pixel value. The monitor was positioned in rooms withstandard indoor lighting levels and minimal natural light.Prey were 39.38 mm wide (126 pixels, approximately4.59°) by 20.03 mm (64 pixels, approx. 2.30°) high, viewedfrom a distance of approximately 500 mm (approx.27.9 pixels per degree). If participants touched the screenwithin the bounds of the prey the code recorded a captureevent to the nearest millisecond, a high-pitched auditorycue sounded, and a green circle appeared around the preyfor 1 s. If they touched the screen outside the prey’sbounds a low-pitched auditory cue sounded, and theywere not progressed to the next screen. If the participantfailed to find the prey after 20 s (timeout) a red circle ap-peared around the moth for 1 s and a low-pitched

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 11 of 13

auditory cue sounded, and capture time was set at 20 s(this occurred in just 3.5% of slides). In addition, for everysuccessful or failed capture event, or timeout event the lo-cation of the touch was recorded. Participants started eachsession by clicking a box asking them to ‘find the artificialtriangular “moths” as fast as possible’, confirming that theywere free to leave at any point, and that it should take lessthan 10 min to complete (all trails were under 10 min). Atotal of 120 participants were tested, each receiving 32slides (i.e. 32 potential capture events), creating a total of3840 unique prey presentations.

StatisticsAll statistics were performed in R version 3.2.2 [50]. Foreach camouflage metric a linear mixed effects modelwas specified using lme4 (version 1.1-10). Thedependent variable in each model was log capture time.The main aim of this study was to establish which cam-ouflage measurement models best predicted human per-formance, and as such we compared the variance incapture times explained between models. The multiplemodels created here increase the likelihood of making atype I error, however, alpha correction methods (such asBonferroni or Šidák corrections) are not strictly suitablefor these data as many of the models are non-independent, measuring subtly different versions of thesame effect, and they would increase the likelihood oftype II errors. As such we focused on the level of vari-ance explained by each variable and its associated effectsizes for ranking the models. A number of variablesknown to affect capture times were included in themodel to reduce the residual variance to be explained bythe camouflage metrics [28]. These were the X and Yscreen coordinates of the prey, each included with quad-ratic functions and with an interaction between them toreflect the fact that prey in the centre of the screen weredetected sooner than those at the edges or corners. Avariable was used to distinguish the first slide from allsubsequent slides, describing whether the participantwas naive to the appearance of the prey. Slide numberwas fitted to account for learning effects. Random effectsfitted to the model were participant ID and backgroundimage ID, allowing the model to ignore the differencesin capture time between participants or against specificbackgrounds when calculating the fixed effects. Eachcamouflage metric was substituted into this model and thedeviance explained by each camouflage metric wascalculated using the pamer function of LMERConvenience-Functions (version 2.10). All camouflage metrics werecontinuous variables transformed where necessary to ex-hibit a normal error distribution with the exception of treat-ment type, which was categorical (background matching ordisruptive prey). An additional final model was assembledbased on the best performing edge disruption metric,

pattern matching metric and luminance matching metricwith two-way interactions between them. These variableswere checked for autocorrelation using Spearman covari-ance matrices [51]. This full model was them simplifiedbased on AIC maximum likelihood model selection todetermine whether the best camouflage predictors interactin synergy to better predict camouflage than any singlemetric on its own.

Additional file

Additional file 1: Table S1. Model terms in the simplified model ofbandpass-based descriptive statistics. Prey X and Y screen coordinates areadded with polynomial fits and an interaction. (DOC 140 kb)

AbbreviationsHMAX: “Hierarchical Model and X”; SIFT: “Scale invariant feature transform”

AcknowledgementsWe thank two anonymous referees for their constructive reviews.

FundingAll authors were funded by a BBSRC grant BB/L017709/1 to MS and JS.

Availability of data and materialsData and scripts are available to download from Open Research Exeter:http://hdl.handle.net/10871/24464. The GabRat measurement tool has alsobeen included in the Multispectral Image Calibration and Analysis Toolboxavailable from http://www.jolyon.co.uk.

Authors’ contributionsJT, JS and MS devised the study. JT created the stimuli and touchscreengame, created and measured the camouflage metrics, performed thestatistical analysis and wrote the first draft of the manuscript. All authorscontributed to the manuscript. All authors read and approved the finalmanuscript.

Competing interestsThe authors declare that they have no competing interests.

Consent for publicationNot applicable.

Ethics approval and consent to participateEthical approval was granted by Exeter University (ID: 2015/736). Participantsconsented to their data being used for scientific purposes by clicking theinitial start screen. The start screen also informed participants that they werefree to leave the experiment at any point. No personal identifying data werecollected.

Author details1University of Exeter, Centre for Ecology and Conservation, College of Life &Environmental Sciences, Penryn Campus, Penryn, Cornwall TR10 9FE, UK.2Centre for Behaviour and Evolution, Institute of Neuroscience, NewcastleUniversity, Henry Wellcome Building, Framlington Place, Newcastle uponTyne NE2 4HH, UK.

Received: 8 October 2016 Accepted: 17 December 2016

References1. Bates HW. Contributions to an Insect Fauna of the Amazon Valley.

Lepidoptera: Heliconid. Trans Linn Soc Lond. 1862;23:495–566.2. Wallace AR. Mimicry and Other Protective Resemblances among Animals.

Volume 88. London: Westminster Review; 1867.3. Kettlewell HBD. Selection experiments on industrial melanism in the

Lepidoptera. Heredity. 1955;9:323–42.

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 12 of 13

4. Cuthill IC, Stevens M, Sheppard J, Maddocks T, Párraga CA, Troscianko TS.Disruptive coloration and background pattern matching. Nature.2005;434:72–4.

5. Endler JA, Mielke PW. Comparing entire colour patterns as birds see them.Biol J Linn Soc. 2005;86:405–31.

6. Skelhorn J, Rowland HM, Speed MP, Ruxton GD. Masquerade: camouflagewithout crypsis. Science. 2010;327:51.

7. Cott HB. Adaptive Coloration in Animals. London: Methuen & Co LTD; 1940.8. Stevens M, Merilaita S. Defining disruptive coloration and distinguishing its

functions. Philos Trans R Soc B Biol Sci. 2009;364:481–8.9. Endler JA. A framework for analysing colour pattern geometry: adjacent

colours. Biol J Linn Soc. 2012;107:233–53.10. Troscianko J, Stevens M. Image Calibration and Analysis Toolbox—a free

software suite for objectively measuring reflectance, colour and pattern.Methods Ecol Evol. 2015;6:1320–31.

11. Renoult JP, Kelber A, Schaefer HM. Colour spaces in ecology andevolutionary biology. In: Biol Rev. 2015. doi:10.1111/brv.12230.

12. Barnett JB, Cuthill IC. Distance-dependent defensive coloration. Curr Biol.2014;24:R1157–8.

13. Stevens M. Predator perception and the interrelation between differentforms of protective coloration. Proc R Soc B Biol Sci. 2007;274:1457–64.

14. Endler JA. A Predator’s View of Animal Color Patterns. In: Hecht MK, SteereWC, Wallace B, editors. In Evolutionary Biology. US: Springer; 1978. p. 319–64[Evolutionary Biology, vol. 11].

15. Thayer AH. Concealing-Coloration in the Animal Kingdom. New York:Macmillan Company; 1909.

16. Troscianko J, Wilson-Aggarwal J, Stevens M, Spottiswoode CN. Camouflagepredicts survival in ground-nesting birds. Sci Rep. 2016;6:19966.

17. Godfrey D, Lythgoe JN, Rumball DA. Zebra stripes and tiger stripes: thespatial frequency distribution of the pattern compared to that of thebackground is significant in display and crypsis. Biol J Linn Soc.1987;32:427–33.

18. Lowe DG. Object recognition from local scale-invariant features. In: InComputer vision, 1999. The proceedings of the seventh IEEE internationalconference on. Volume 2. Ieee. 1999. p. 1150–7.

19. Stevens M, Cuthill IC. Disruptive coloration, crypsis and edge detection inearly visual processing. Proc R Soc B Biol Sci. 2006;273:2141–7.

20. Serre T, Wolf L, Bileschi S, Riesenhuber M, Poggio T. Robust objectrecognition with cortex-like mechanisms. Pattern Anal Mach Intell IEEE TransOn. 2007;29:411–26.

21. Lovell PG, Ruxton GD, Langridge KV, Spencer KA. Egg-laying substrateselection for optimal camouflage by quail. Curr Biol CB. 2013;23:260–4.

22. Stoddard MC, Kilner RM, Town C. Pattern recognition algorithm reveals howbirds evolve individual egg pattern signatures. Nat Commun. 2014;5:4117.

23. Cook LM, Grant BS, Saccheri IJ, Mallet J. Selective bird predation on thepeppered moth: the last experiment of Michael Majerus. Biol Lett.2012;8:609–12.

24. Xiao F, Cuthill IC. Background complexity and the detectability ofcamouflaged targets by birds and humans. Proc R Soc B. 2016;283:20161527.

25. Merilaita S, Lind J. Background-matching and disruptive coloration, and theevolution of cryptic coloration. Proc R Soc Lond B Biol Sci. 2005;272:665–70.

26. Stevens M, Winney IS, Cantor A, Graham J. Outline and surface disruption inanimal camouflage. Proc R Soc B Biol Sci. 2009;276:781–6.

27. Stevens M, Marshall KL, Troscianko J, Finlay S, Burnand D, Chadwick SL.Revealed by conspicuousness: distractive markings reduce camouflage.Behav Ecol. 2013;24:213–22.

28. Troscianko J, Lown AE, Hughes AE, Stevens M. Defeating crypsis: detectionand learning of camouflage strategies. PLoS One. 2013;8:e73733.

29. Espinosa I, Cuthill IC. Disruptive Colouration and Perceptual Grouping. PLoSONE. 2014;9:e87153.

30. Webster RJ, Godin J-GJ, Sherratt TN. The role of body shape and edgecharacteristics on the concealment afforded by potentially disruptivemarking. Anim Behav. 2015;104:197–202.

31. Osorio D, Srinivasan MV. Camouflage by edge enhancement in animalcoloration patterns and its implications for visual mechanisms. Proc R SocLond B Biol Sci. 1991;244:81–5.

32. Webster RJ, Hassall C, Herdman CM, Godin J-GJ, Sherratt TN. Disruptivecamouflage impairs object recognition. Biol Lett. 2013;9:20130501.

33. Kang C, Stevens M, Moon J, Lee S-I, Jablonski PG. Camouflage throughbehavior in moths: the role of background matching and disruptivecoloration. Behav Ecol. 2014;26:45–54.

34. Troscianko T, Benton CP, Lovell PG, Tolhurst DJ, Pizlo Z. Camouflage andvisual perception. Philos Trans R Soc B Biol Sci. 2009;364:449–61.

35. Marĉelja S. Mathematical description of the responses of simple corticalcells*. JOSA. 1980;70:1297–300.

36. Jones JP, Palmer LA. An evaluation of the two-dimensional Gabor filtermodel of simple receptive fields in cat striate cortex. J Neurophysiol. 1987;58:1233–58.

37. Gabor D. Theory of communication. Part 1: The analysis of information.Electr Eng-Part III Radio Commun Eng J Inst Of. 1946;93:429–41.

38. Rosenholtz R, Li Y, Mansfield J, Jin Z. Feature congestion: a measure ofdisplay clutter. In: In Proceedings of the SIGCHI conference on Humanfactors in computing systems. New York: ACM; 2005. p. 761–70.

39. King-Smith PE, Kulikowski JJ. The detection of gratings by independentactivation of line detectors. J Physiol. 1975;247:237–71.

40. Chiao C-C, Chubb C, Buresch K, Siemann L, Hanlon RT. The scaling effects ofsubstrate texture on camouflage patterning in cuttlefish. Vision Res.2009;49:1647–56.

41. Spottiswoode CN, Stevens M. Visual modeling shows that avian hostparents use multiple visual cues in rejecting parasitic eggs. Proc Natl AcadSci. 2010;107:8672–6.

42. Moreno P, Marín-Jiménez MJ, Bernardino A, Santos-Victor J, de la NP B. AComparative Study of Local Descriptors for Object Category Recognition:SIFT vs HMAX. In: Martí J, Benedí JM, Mendonça AM, Serrat J, editors. InPattern Recognition and Image Analysis. Berlin, Heidelberg: Springer; 2007.p. 515–22 [Lecture Notes in Computer Science, vol. 4477].

43. Dimitrova M, Merilaita S. Hide and seek: properties of prey and backgroundpatterns affect prey detection by blue tits. Behav Ecol. 2014;25:402–8.

44. Livingstone M, Hubel D. Segregation of form, color, movement, and depth:anatomy, physiology, and perception. Science. 1988;240:740–9.

45. Riesenhuber M, Poggio T. Hierarchical models of object recognition incortex. Nat Neurosci. 1999;2:1019–25.

46. Serre T, Kouh M, Cadieu C, Knoblich U, Kreiman G, Poggio T. A Theory ofObject Recognition: Computations and Circuits in the Feedforward Path ofthe Ventral Stream in Primate Visual Cortex. 2005. DTIC Document.

47. Schindelin J, Arganda-Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T,Preibisch S, Rueden C, Saalfeld S, Schmid B, et al. Fiji: an open-sourceplatform for biological-image analysis. Nat Methods. 2012;9:676–82.

48. Spottiswoode CN, Stevens M. How to evade a coevolving brood parasite:egg discrimination versus egg variability as host defences. Proc R Soc LondB Biol Sci. 2011;278:3566–73.

49. Gibara T. Canny edge detection algoritm in Java. 2011.50. R Core Team. R: A Language and Environment for Statistical Computing.

Vienna: R Foundation for Statistical Computing; 2013.51. Zuur A, Ieno EN, Walker N, Saveliev AA, Smith GM. Mixed Effects Models

and Extensions in Ecology with R. New York: Springer; 2009.

• We accept pre-submission inquiries

• Our selector tool helps you to find the most relevant journal

• We provide round the clock customer support

• Convenient online submission

• Thorough peer review

• Inclusion in PubMed and all major indexing services

• Maximum visibility for your research

Submit your manuscript atwww.biomedcentral.com/submit

Submit your next manuscript to BioMed Central and we will help you at every step:

Troscianko et al. BMC Evolutionary Biology (2017) 17:7 Page 13 of 13