Image Denoising using new digital pulse mode neural network

7/27/2019 Image Denoising using new digital pulse mode neural network

http://slidepdf.com/reader/full/image-denoising-using-new-digital-pulse-mode-neural-network 1/5

I nternational Jou rn al of Engin eeri ng Trends and Technology- Volu me4I ssue2- 2013

ISSN: 2231-5381 http://www.internationaljournalssrg.org Page 114

Image Denoising using new digital pulse mode neural network

Suresh .R (ME) 1 , Kannadhasan .G(ME) 2, Jebin.P.L(ME) 3

Embedded System technologies, S A Engineering College,Chennai-77, Tamil nadu, India.

Abstract — I n th is paper, we propose a new architecture of Pul se M ode Neural N etwork (PM NN) wi th very simple activation function. ANNs are usually im plemented in softwar e, but many appli cations requir e impl ementation of f ast and l arge neur al n etworks on effi cient custom device. Pulse mode is gain in g suppor t in the field of hardware Neural Networks thanks to its hi gher density of in tegration. H ardware implementation is thu s devoted onl y to the general ization ph ase using th e network par ameters set. In thi s context, the main i dea is to apply a new kind of activation function, simply generated by the product of two sigmoidal functions,which are very simple and already implemented in previous work. The pulse mode ori gin ates fr om where a stochastic computin g is applied using basic logi c gates and random pulse sequences as inputs. The filtered results are verif ied in term s of the Peak Signal to Noise Ratio (PSNR). Experimental results reveal that the proposed PM NN fi lter h as a greater abili ty to recover the in formati ve pixel intensities fr om the inf ected image with a recovery of 7.5 dB for Gaussian noi se and 5.3 dB for Speckle noise. I n thi s work, l ooking f or simple implementation and efficient learning, we propose to impl ement a new kind of adjustable highl y modulator fun ction based on simple existing non li near sigmoidal blocks. Besides, such results demonstrate the

performance and efficiency of our Neural filter when compared to other conventional fi lterin g techni ques.

Key words : PSNR, PMN N, Sigmoidal,Gaussian noise,fi lter.

1. INTRODUCTIONImage processing is an important component

of modern technologies because human depends somuch on the visual information than other creatures.Image is better than any other information form for usto perceive. Among our information about the world,99% is perceived with our eyes . Image processing hastraditionally been an area in the engineering

community.Today, the medical industry, astronomy,

physics, chemistry, forensics, remote sensing,manufacturing, and defense are just some of the manyfields that rely upon images to store, display, and

provide information about the world around us. Thechallenge to scientists, engineers and business people isto quickly extract valuable information from raw imagedata. This is the primary purpose of image processing – converting images to information.

2.PROJECT DESCRIPTIONImage denoising is an important image

processing task, both as a process itself, and as acomponent in other processes. Very many ways todenoise an image or a set of data exists. The main

properties of a good image denoising model is that itwill remove noise while preserving edges.Traditionally, linear models have been used. Onecommon approach is to use a Gaussian filter, or

equivalently solving the heat equation with the noisyimage as input data. In some cases, the activationvalues of the units undergo a relaxation process suchthat the network will evolve to a stable state in whichthese activations do not change anymore.

One big advantage of linear noise removalmodels is the speed. The linear models is that theyare able to preserve edges in a good manner. Edges,which are recognized as discontinuities in the image..

Nonlinear models on the other hand can handle edgesin a much better way than linear models . One

popular model for nonlinear image denoising is themodified median filter. This filter is very good at

preserving edges, but smoothly varying regions in theinput image are transformed into piecewise constantregions in the output image.

Block diagramThe basic architecture consists of three types

of neuron layers: input, hidden, and output. In feed-forward networks, the signal flow is from input tooutput units, strictly in a feed-forward direction. Thedata processing can extend over multiple layers of units, but no feedback connections are present.

2.1. Input mask

http://www.internationaljournalssrg.org/







Masking is available for selectedenvironmental functions including statistics,classification, unmixing, matched filtering,continuum removal and spectral feature fitting. Use

masking to create image masks.Selecting Areas : Select from the following options to define mask areas

To set the defined areas in the mask to 1(On) or to 0 (Off).The mask is built using aLogical OR or Logical AND operation

between all of the items in the list. To define the mask using only those areas

where the listed data ranges, annotationshapes and select Logical AND.

To use all the defined areas to make themask, select Logical OR.

2.2.Input LayerEvery neural network has the number of

neurons comprising this layer, this parameter iscompletely and uniquely determined once know theshape of the training data. Specifically, the number of neurons comprising that layer is equal to thenumber of features (columns) of the data. Someneural network configurations add one additionalnode for a bias term.

2.3 Hidden LayerThe hidden layer configuration using just two rules:

(i) Number of hidden layers equals one(ii) The number of neurons in that layer is

the mean of the neurons in the input andoutput layers.

A set of hidden units which take input fromthe input layer. The hidden units collectively form thehidden layer. For simplicity, we assume that eachunit in the input layer is connected to each unit of thehidden layer, but this isn't necessarily the case. Aweighted sum of the output from the input unitsforms the input to every hidden unit. Note that thenumber of hidden units is usually smaller than thenumber of input units.

Assigning weights in the hidden layer

In order to assigning the weight able to perform our categorisation task, we need to use theexamples to train the weights between the input unitsand the output unit, and to train the threshold. Tosimplify the routine, we think of the threshold as aspecial weight, which comes from a special inputnode that always outputs a 1.

S

The output from the perception is +1 if theweighted sum from all the input units (including thespecial one) is greater than zero, and it outputs -1otherwise. The assigning weight w0 is simply thethreshold value. However, thinking of the network like this means we can train w0 in the same way aswe train all the other weights. Then, given that thenetwork should have calculated the target value t(E)for example E, but actually calculated the observedvalue o(E), then Δ is calculated as: Δ = η (t(E) -o(E))x i

2.4 Output LayerThe output layer like the Input layer, every

neural network has exactly one output layer.Determining its size (number of neurons) is simple. Itis completely determined by the chosen modelconfiguration.

Neural network is running in Machine Mode or Regression Mode (the ML convention of using a term that is also used in statistics butassigning a different meaning to it is very confusing).If the neural network is a regressor, then the outputlayer has a single node.

3. NEURAL NETWORK SIGMOIDALFUNCTION

A sigmoid function is a boundeddifferentiable real function that is defined for all real








input values and that has a positive derivativeeverywhere.

The key point is that this architecture isvery simple and very generalized. This same flow

diagram can be used for many problems, regardlessof their particular process. The ability of the neuralnetwork to provide useful data manipulation lies inthe proper selection of the weights. This is adramatic departure from conventional information

processing where solutions are described in step bystep procedures.

As an example, imagine a neural network for recognizing objects in a sonars ignal. Supposethat 1000 samples from the signal are stored in acomputer. How does the computer determine if thesedata represent a submarine, whale,underseamountain, or nothing at all? Conventional DSPwould approach this problem with mathematics andalgorithms, such as correlation and frequencyspectrum analysis. With a neural network, the 1000samples are simply fed intothe input layer, resultingin values popping from the output layer. By selectingthe proper weights, the output can be configured toreport a wide range ofinformation. For instance,there might be outputs for: submarine (yes/no),whale(yes/no), undersea mountain (yes/no), etc.

Neural network sigmoidal function

The sigmoid activation function is thedefault choice for the Feed forward Layer class. It is

possible to use others. For example, to use thehyperbolic tangent activation function, the following

lines of code would be used to create the layers.

network.addLayer(new FeedforwardLayer(newActivation(),2));network.addLayer(new FeedforwardLayer(newActivation(),3));network.addLayer(new FeedforwardLayer(newActivation(),1));

A sigmoid activation function uses thesigmoid function to determine its activation. Thesigmoid function is defined as follows:

( )

Sigmoid function

4. DENOISING THE IMAGE

Noise is the result of errors in the imageacquisition process that result in pixel values that donot reflect the true intensities of the real scene. Thereare several ways that noise can be introduced into animage, depending on how the image is created.

The different filter as a denoiser leads tosolving a 2nd order nonlinear PDE. Since smoothregions are transformed into piecewise constantregions when using the filter, it is desirable to createa model for which smoothly varying regions aretransformed into smoothly varying regions, and yetthe edges are preserved. The median is much lesssensitive than the mean to extreme values (calledoutliers).

This can be done for instance by solving a4th order PDEd instead of the 2nd order PDE fromthe filter. Results show that the 4th order filter

produces much better results in smooth regions, andstill preserves edges in a very good way.

Original image








Noisy image

Restored image

True normal image

Noisy normal image

Smoothed normal image

The three images above show a smallexcerpt of the normal vectors of the above shownimage. The first image shows the normal of theoriginal image, the middle image shows the normalof the noisy image, and the last image shows the

smoothed normal.

4.1. Different Types of Noises

Gaussian noiseGaussian noise is evenly distributed over the

signal . This means that each pixel in the noisy imageis the sum of the true pixel value and a randomGaussian distribute noise value. As the nameindicates, this type of noise has a Gaussiandistribution, which has a bell shaped probabilitydistribution function.

Salt and Pepper NoiseSalt and pepper noise [Um98] is an impulsetype of noise, which is also referred to as intensityspikes. This is caused generally due to errors in datatransmission. It has only two possible values, a and b.

The probability of each is typically less than0.1. The corrupted pixels are set alternatively to theminimum or to the maximum value, giving the imagea “salt and pepper” like appearance. Unaffected

pixels remain unchanged. For an 8-bit image, thetypical value for pepper noise is 0 and for salt noise255. The salt and pepper noise is generally caused bymalfunctioning of pixel elements in the camerasensors, faulty memory locations, or timing errors inthe digitization process.

Speckle Noise Speckle noise is a multiplicative noise. This

type of noise occurs in almost all coherent imagingsystems such as laser, acoustics and SAR(SyntheticAperture Radar) imagery. The source of this noise isattributed to random interference between the








coherent returns. Fully developed speckle noise hasthe characteristic of multiplicative noise.

Brownian Noise

Brownian noise comes under the categoryof fractal or 1/ f noises. The mathematical model for 1/ f noise is fractional Brownian motion . FractalBrownian motion is a non-stationary stochastic

process that follows a normal distribution. Browniannoise is a special case of 1/ f noise. It is obtained byintegrating white noise

5. CONCLUSION AND FUTUREENHANCEMENTS

The whole network was implemented incomputer and Microcontorller chip. Suchimplementation offers many advantages over other solution with respect to both hardwareimplementation cost and device timing performance.Denoising of an image reduces significantly theamount of data and filters out information that may

be regarded as less irrelevant. Denoising is efficientin medical imaging. By using floating point number system for synapse weight value representation, anyfunction can be approximated by the network.

The most important feature of the proposedfunction is the simplicity of implementation and theeasy of programmability, not making use of anycomplex explicit function to be implemented, which

increase the learning capacity and enhance the Network efficiency.

6. REFERENCES

1. Amir Gargouri, Mohamed Krid, DorraSellami Masmoudi , New Digital Pulse -Mode Neural Network based ImageDenoising” 2012

2. Hikawa.H, „A digital hardware pulse-modeneuron with piecewise Linear activationfunction, IEEE Transactions on Neural

Networks , vol. 14 pp. 1028– 1037, 2003.

3. Kaliraj.G, and Baskar.S , „An efficientapproach for the removal of impulse noisefrom the corrupted image using neuralnetwork based impulse detector, Image andVision Computing, vol. 28, pp. 458 – 466,2010.

4. Krid.M, Damak.A, and Masmoudi.D.S ,„Hardware implementation of a pulse modeneural network-based edge detection system,AEU International Journal of Electronics

and Communications , vol. 63, no. 10, pp.810-820, 2009.

5. Lee.C.Y and Lin.C.J, „Implementation of a neuro-fuzzy network with on-chip learningand its applications , Expert Systems withApplications, vol. 38, pp. 673-681, 2011.

6. Lina. C. J and Tsaib.H.M, „FPGAimplementation of a wavelet neural network with particle swarm optimizat ion Learning ,Mathematical and Computer Modelling, vol.47, pp. 982-996, 2008.




Image Denoising using new digital pulse mode neural network

Documents