Imran Sarwar Bajwa, S. Irfan Hyder [2005], "PCA Based Image Classification of Single-Layered Cloud Types", in 1st IEEE International Conference on Emerging Technologies (ICET 2005), Islamabad, Pakistan, Jan 2005, pp:365-369
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
IEEE --- 2005 Interntational Conference on Emnerging TechnologiesSeptember 17-18, Islamiiabad
PCA based Image Classification of Single-layered CloudTypes
Imran Sarwar Bajwa, Syed Irfan HyderPAF- Karachi Institute ofEconomics and Technology
imransbiawa(divahoo.com,nhyder(pafkiet. edu.pk
AbstractThe paper presents an automatic classification
system, which discriminates the diferent types of'sinigle-layered clouds uising Principal Component Anialysis(PCA) with enhanced accuracy as compared to othertechniques. PCA is an image classification techniqutetypically used for ace recognition. Principalcomponents are the distinctive or peculiar featuires of animage. The approach described in this paper iuses thisPCA capability Jbr enhancing the accuracy of cloudimage analysis. To demonstrate this enhancement, asoJtware classifier system has been developed thatincorporates PCA capability Jbr better discrimination ofclouid images. The system is first trained using clouidimages. In training phase, system reads mqjor principalfeatures of the different cloud images to produce animage space. In testing phase, a new cloud image can beclassified by comparing it with the specified image spaceusing the PCA algorithm.
Keywords: PCA, single cloud type recognition,principal components, eigenvectors, weather prediction,image recognition.
1. Introduction
Weather forecasting applications use various patternrecognition techniques to analyze clouds' information andother meteorological parameters. Neural Networks is anoften-used approach [3], [13] for image processing. Somestatistical methodologies like FDA [4], RBFNN [I] andSVM [12] are also being used for image analysis. Thesemethodologies require more training time and havelimited accuracy of about 70% [ 11]. This level ofaccuracy often degrades classification of clouds, andhence the accuracy of rain and other weather predictionsis reduced [15].
Principal Component Analysis (PCA) treats each imageas an entity. A set of images is needed, which defines aclass based on the core feature, derived from thoseimages [6]. A single class can cover a certain number ofimages. The number of images in a class can beineffective but the image quality can invariantly impactthe overall image analysis results and consequences. PCAis used to avoid computational intense calculations. Itsuse results in fast and relatively more accurate inferences
[7]. PCA is a way of identifying patterns in data andexpressing the data in a way that highlights its similaritiesand differences. PCA strives to identify relatively fewer"features" or components that as a whole represent thefull object state and hence are appropriately termed"Principal Components". Thus, principal componentsextracted by PCA implicitly represent all the features.However, these abstracted features may or may notinclude a specific feature [5].
Better accuracy in cloud classification meansaccurate categorization of clouds according to high, midand low levels. These high, mid and low-level clouds arefurther classified in their particular sub classes illustratedin Section 3.3. PCA can easily handle a large amount ofdata due to its capability of reducing data dimensionalityand complexity, thus getting better results [4]. Thealgorithm provides a more accurate cloud classificationthat yield better and concise forecasting of rain.
1.1 PCA and EigenvectorsTraining procedure in any classification system is
significant and can be beneficial. Using variousalgorithms training can be performed. Algorithm used incurrent image classifier system for training is principalcomponent analysis, whose major emphasis is to locateand depict the principal features of the given sampleimage [12].
The Eigenvectors is a key capability used in PCAanalysis algorithm. Eigenvectors are defined to be arelated set of spatial characteristics [6] that computer usesto recognize a specific cloud type. PCA technique usestraining and testing sets of images. Eigenvectors of thecovariance matrix is computed from the training set ofimages. These eigenvectors represent the principalcomponents of the training images [7]. Theseeigenvectors are often ortho-normal to each other. In thecontext of clouds classification, these eigenvectors wouldform the cloud space. They may not correspond directlyto any cloud feature like height, width and density. Whenthe eigenvectors are displayed, they look like a ghostlycloud. They can be thought of as a set of features thattogether characterize the variation between cloud images.
Cloud Detection consists in locating a cloud incomplex scenery, by locating and cutting it out. Somemethods search elliptical and polygonal forms [2], others
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
IEEE --- 2005 International Conference on Emnerging TechnologiesSeptemtber 17-18, Islaminabad
seek the texture and color of the clouds and still othersseek the patterns and boundaries of the cloud [3].
PCA is used abundantly for image analysis andclassification purposes [6] because it is a simple, non-parametric method of extracting relevant informationfrom confusing data sets [1]. PCA extracts relevantfeatures from data by performing an orthogonaltransformation. PCA also provides assistance to reduce acomplex data set to a lower dimension [7]. This paperdemonstrates that PCA is relatively better than othertechniques in discriminating different types of single-layered cloud images.
2. Methodology
2.1 System's General ArchitectureThe developed classifier system discriminates the
single-layered cloud types. It carries out classification infive modules: image acquisition, detection, extraction ofrelated attributes, comparison of these attributes andfinally classification as shown in Fig. 1.
Cloud Extraction |Imag Detection of Attributes
Cloudi Type I1Inage < /ass Classificatio Comparison
Fig 1: General architecture of the cloud types classificationsystem
Image acquirerThis module helps to acquire new image. The image
can be acquired through different sources e.g. digitalcamera. Images for testing and training phases areconverted to 256-bit Gray color image. Images are alsoscaled to 50 x 50 ratio. This ratio can vary from 30 x 30to 50 x 50.
Cloued DetectionThis module detects the presence of the cloud
fragments in the images. In this module it is specifiedthat rather the images contain the clouds or not.
Extraction of AttributesThis module identifies the various patterns in data.
Cloud attributes are extracted from images usingPrincipal Component Analysis algorithm.
lmage ComparisonThis module compares the principal features of the
test image with the image-space of already given images
in training set. After matching it infers that rather imageis recognized or not.
Clouid Tvpe ClassificationThis module finally detects and classifies the cloud
type. Images are classified to their respective types on thebasis of the matching inferences provided by previousmodule.
2.2 PCA based Extraction of the AttributesThe system presented in this work exemplifies the
concept of Eigenvectors. These eigenvectors are a smallgroup of characteristics extracted by the designedclassifier system using PCA. PCA is a two-phasealgorithm.
* Training* Recognition
2.2.1 TrainingTraining phase constructs an image-space, called a
cloud space, which is later required for classification intesting phase. In training phase, the classifier system istrained by using sample data input. If it is required, outputpattem can be enhanced and improvised by retraining thesystem by more refined and conspicuous data. Training isperformed using n images in the following 6 steps:
STEP 1:Each sample image is converted into a row vector. A rowvector can be constructed by concatenating each row withfirst row in sequence. As in fig-2 a m x n matrix isconverted into a single row 1 x mn vector X.
X~~~~~~~~~~Xx~~~~~~~~~~~x
Fig 2: A row vector representation of a 2-D cloud image
STEP 2The row vector matrix is constructed by combining
together the row vectors of n cloud images. Xi is a rowvector of a sample image i, where i = 1 .. n.
\~~~~~~~~~~~~~~......41J1 + . *
Fi.:Woecoddsrbto.Rwvco fX
366
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
IEEE --- 2005 International Conference on Emnerging TechnologiesSeptemlber 17-18, Islamabad
STEP 3
A mean cloud vector T of n row vectors is calculatedto extract required principal features.
VJ'= (/)n) EX, Where i = I .. n
STEP 4
A new matrix ( is constructed by subtracting mean cloudvector ¶1 from each cloud image X of the training set.
<P =XjI
STEP 5A data covariance matrix C is calculated by multiplyingmatrix ( with its transpose matrix 0,.
C=- , (
STEP 6
Twenty highest valued eigenvectors are then pickedto make an image space from the resultant covariancematrix C.
*(StepS 6 are performed using Matlab 5.6)
2.2.2 RecognitionIn testing phase, each new image is analyzed and its
principal features are located. Then these principalfeatures are compared with the principal features ofimage-space. If some match is found there, then the imageis classified according to the previously defined rules.Recognition or testing phase is performed in thefollowing two steps.
STEP I
A new cloud image is categorized by calculatingprojection Q on image-space by
n=U * (Z - )Where Ui is image-space and Z is the new Image
STEP 2
If threshold (P matches with one of the thresholds inimage space then cloud recognition occurs and theparticular cloud type is specified.
,(i = I/k max (Qi -2) where (i,j= ..... n)
3. Experiments
A series of experiments were done using thedeveloped classified system to evaluate its accuracy.Experiments were performed using following steps:
1- Data Collection2- Normalizing Cloud Images3- Define Classes
4- Cloud Type Classification5- Evaluate Accuracy6- Comparison with other Technologies
3.1 Data CollectionImage data of cloud's different types was obtained
for training purposes. This data is available from differentsources. Ground-based cloud images have been used inthis experiment. These images of the general and sub-cloud types are available at different websites of world'smajor weather forecasting organizations [13], [14] , [15].Overlapping sets of training images and testing imageswere used for the experiment. Global daytime cloudimages are used in development and implementationaspects of a principal component analysis classificationsystem. The designed image classifier system is used tofind the presence of clouds and classification of single-layer clouds in cloud images.
3.2 Normalizing Cloud ImagesImages for testing and training phase are of 256-bit
Gray color image and are overlapping. If the acquiredimage is not in specified bitmap format then it isconverted into required format. The system obtains theimage in the form ofBMP of JPEG format.
Acquired Image was of size 50 x 50 pixels forprocessing in the designed system. But this ratio can betuned from 30 x 30 to 50 x 50. This module gets theimage in integer or short co-ordinate i.e. perform scalingat 50 x 50 scale.
3.3 Define ClassesAn efficient and effective image classifier system oftenconsists of a defined set of classes. These preciselydefined classes are well separated by a set of features thatare typically derived from the multi-dimensionalradiometric image data. The selection of classes is ofteninfluenced by desired application and classes may becomplicated. In this research, there are four definedgeneral classes and these general classes are furtherdivided into sub classes and they are
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
IEEE --- 2005 International Conference on Emiierging TechnologiesSeptemnber 17-18, Islanmabad
3.4 Cloud Type ClassificationTwo types of satellite images have been used, first as
training image and other images with clouds for testing.Comparing the individual pixel values within 50 x 50array with a clear sky images depicts cloud fragments in asample image. Often the array of 32 x 32 array is used inconventional image recognition applications. As thegreater number of pixels can immensely affect thememory usage so array of smaller range is preferred. Butthis procedure also affects the overall image processingaccuracy. But PCA handles images so conveniently thatan array of greater range may be used to get still higheraccuracy.
If the cloud matches with the existing collected datathen the program will display as match is found. Itdisplays cloud's general type as low, mid or high.Program has also the capacity of prescribing the cloud'ssub type and also describing the properties of each cloudsub type.
3.5 Evaluate AccuracyTo test the accuracy of the designed system images
with clear sky and images with all types of single-layeredclouds are used. 17 Clear sky images were used fortesting and all images were successfully categorized. 36Images of single-layered cloud types were used andshowed results with high accuracy. A matrix of results oftesting images is shown below.
Table 1. Testing results of different cloud type images
A matrix representing classification accuracy test (%)for cloud free and single-layered cloud types isconstructed. Classification inferences of PrincipalComponent Analysis for different cloud types are shownin the matrix. Overall classification accuracy for single-layered clouds is determined by dividing number ofcorrectly classified samples by the total number ofsamples. An accuracy test (%) table is shown here.
3.6 Comparison with other TechniquesVarious classification techniques and algorithms are
used for image classification. Each technique has its ownrespective accuracy level. The derived results usingprincipal component analysis are compared with theresults of other technologies used for cloud classification.Different technologies provide with different levelaccuracies.
Results show that PCA, relative to other statisticaltechniques, is more accurate [table. 3]. Other statisticaltechniques include Fuzzy Logic based systems that give84% accuracy [5] but Fuzzy systems are dependent on theappropriateness of the initial categories defined i.e. mucheffort is needed for domain knowledge and efficiencyissues. Neural networks demand intense domainknowledge and intuition for representation otherwisesuffer from divergent training sessions and inaccurateresults [11].
PCA is the image classification technique, whichprovides higher accuracy up to 90%. Statistics show thatPCA based image classifier system is a better classifierthan other used techniques. A comparison of PCA withother techniques is also given below.
3.6.1.1.1 Technology Name Accuracy Error_______ ______ ______ ______ ______ ______ Per. Ratio
Table 3. Accuracy comparison in different techniques
4. Conclusion & Future Work
PCA is an efficient identifier in terms of time andprovides better accuracy in cloud image recognition. APCA-based system provides high speed processing withrelatively better accuracy. PCA also easily handles a largeamount of data due to its capability of reducing datadimensionality and complexity. PCA algorithm provides amore accurate cloud classification that infers better and
368
Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
IEEE --- 2005 International Conference on Emnerging TechnologiesSeptember 17-18, Islamabad
concise forecasting of rain. Probably, the more long-termweather forecasting is also possible.
In this report only one type of cloud has beenaddressed and that is single-layered clouds. Other type isMulti-layered clouds. Multi-layered clouds are 60 % to65% of total clouds. So their identification andclassification is also a significant task.
Future goal of the research is to analyze all cloudtypes using satellite image. Satellite image containsclouds of all types and there is very complex infonnation.For simple cloud recognition, co-variance matrix has beenused. To cover the whole scenario (all cloud types)another type of matrix named, co-exist matrix can beused. There is also need of generalizing the algorithm.This process may need much more effort. Furthermore,the accuracy estimations can be further improved using amore carefully selected disjoint sets of image data for thetraining set and the testing set.
5. References
[1 ] Su Hongtao, David Dagan Fetig, Zhao Rong-chun, 1997.Face Recognition Using Multi-feature and Radial BasisFunction Network.
[2] Bankert, R. L., 1994: Cloud classification ofAVHRRimagery in maritime regions using a probabilistic neuralnetwork. Jotirnal ofApplied Meteorology., 33, 909-918.Boulder, CO 80307.
[3] Barsi, A., Heipke, C., Willrich, F., 2002, Detecting roadjunctions by Artificial Neural Networks - JEANS,International Archives oqfPhotogrammetrv. Remote Sensinsgand Spatial Information Science (34) 3B, pp. 18-21
[41 Seong-Wook Joo, December 2003. Face Recognition usingPCA and FDA with intensity normalization.
[5] Bryan A. Baum, Vasanth Tovinkere & Jay Titlow, RonaldM. Welch, 1997. Automated Cloud Classification of GlobalAVHRR Data Using a Fuzzy Logic Approach. Journal ofApplied Meteorology. pp 1519-1539.
[6] Chi-Fa Chen, Yu-Shan Tseng and Chia-Yen Chen, 2003.Combination ofPCA and Wavelet Transforms for FaceRecognition on 2.5D Images. Conf of Image and VisionComputing'03 26-28 November 2003
[7] Cristina Conde ,Antonio Ruiz and Enrique Cabello, 2003.PCA vs Low Resolution Images in Face Verification.Proceedings of the 12th International Conference on ImageAnalysis and Processing (ICIAP'03).
[8] David Guillamet, Bemt Schiele, and Jordi Vitri. AnalyzingNon-Negative Matrix Factorization for Image Classification.Conference on Pattern Recognition ICPR 2002, Quebec,Canada, August 2002
[9] Kishor Saitwal, r. Azimi Sadjadi and Donald Rinki, 1997. Amulti-channel temporarily adaptive system for continuouscloud classification from Satellite Imagery.
[10] Bin Tian, Mamood R. AzimiSadjadi, Thomas H. Vonderand Donald Rienki, 1999. Neural Network based cloudclassification on satellite Imagery using textural features
[11] Simon Haykin, McMaster University, OntarioCanada 1998, Neural Networks a Comprehensive Foundation,2/E, Prentice Hall publishers
[12] D. Cao, 0. Masoud, D. Boley, N. Papanikolopoulos, 2004.Online Motion Classification using Support Vector Machine,IEEE 2004 International Conference on Robotics andAutomation, April 26 - May 1.
[13] VanderZwaag, B.J., Slump, C.H. and Spaanenburg, L.(2002) Analysis of neural networks for edge detection,Proceedings. Pro-Risc'02 (Veldhoven) pp. 580-586