Comparison of medical image compression using DWT algorithm and neural network techniques.
INTRODUCTIONAn image consists of bulky amount of data and requires additional space in the memory. If more number of data is required for transmission then it takes much time to deliver the data to the receiver. Generally, Digital image broadcasting needs a huge amount of data and the time taken for the transmission over communication channels is very high. To overcome this difficulty, a many techniques to compress the amount of data for representing a digital image have been formulated to make its storage and transmission cost-effective. Thus by using image compression techniques the time consumption can be greatly reduced. In this method, the elimination of redundant data in an image can be possible. The image which is compressed occupies less memory space and very less time to broadcast in the form of sequence from transmitter to receiver. The compression which is done in file size permits more images to be stored in the available amount of memory space. It can be greatly used to decreases the time needed for images to be uploaded over the Internet or downloaded from it. Compression means to make file size smaller by reorganizing the data in the file. Image compression adjusts the data and may degrade it to achieve desired compression level, depending on the compression ratio. If there is better compression ratio, the smaller the file size here more data is packed into smaller space, but lower the quality of the compressed product. A digital image which can be obtained by sampling and quantizing a continuous tone picture requires large number of storage area. For example, if we want to send an image with 512*512 pixels it will occupy 1GB storage on a particular disk and also if we want to add two or more images of same size in that same disk it will not be able to fit in a single disk. Thus, in order to transmit such an image over 32.8 Kbps modem would almost take 6-10 minutes. The main purpose for this image compression is to reduce the quantity of data that is essential for representing the sampled digital images and hence reduce the cost for storage space and communication.
Figure 1 explains the block diagram of Image Compression. First we need transform the input Image using Forward Transform and again the quantization of input image is done and then following the Entropy encoding and finally we are getting the Compressed Image. Now we can store or transmit the compressed Image. These are the steps followed in Compression Techniques. Here, the Image Compression Techniques is divided into two types namely Lossy and Lossless Techniques. In lossy, some information is lost during compression of Image whereas in lossless compression no information is lost during Image Compression. Discrete Wavelet Transform (DWT) comes under Lossless Compression. Neural Network training algorithm can be widely divided into Supervised and Unsupervised learning. This Back propagation Neural Networks (BPNN) and Radial Basis Function Neural Networks (RBFNN) comes under Supervised Learning Algorithm.
The remaining part of the paper is organized as follows: Section II involves the detailed description of Compression algorithms and three different types of compression techniques that include DWT algorithm, Back propagation Neural Networks (BPNN) and Radial Basis Function Neural Networks (RBFNN). Section III involves the works related to the detailed description of above three techniques. Section IV involves the performance analysis of different parameters and the results obtained using compression algorithms. Section V involves comparison of the all the three algorithms by providing graph for Compression Ratio (CR) and Peak Signal to Noise Ratio (PSNR). The paper is concluded in Section VI.
MATERIALS AND METHODS
Need For Compression:
As already mentioned, in order to store a color image of a moderate size, e.g. 512*512 pixels it will occupy 1GB storage on a particular disk. Thus, in order to transmit such an image over 32.8 Kbps modem would almost take 6-10 minutes. To store these images, and make them available over network (e.g. the internet), compression techniques are needed. Image compression can be used in reducing the amount of data required to represent a digital image. The basic concept in the reduction process is the removal of redundant data. At receiver, the compressed image is decompressed to reconstruct the original image or an approximation to it.
Principle behind Compression:
Generally, the principle behind compression consists of three types of redundancies. They are:
a. Coding redundancy: Fewer bits to represent frequently occurring symbols.
b. Inter pixel redundancy: Neighboring pixels have almost same value.
c. Psycho visual redundancy: Human visual system cannot be able to distinguish all colors.
Discrete Wavelet Transform:
DWT plays an important role to compress the given image without the loss of any information in that particular image. DWT comes under lossless type of image compression. Here, DWT can be mainly used in the transformation of a discrete time signal to Discrete Wavelet Representation. DWT usually based on time-scale representation, which can be able to provide multi-resolution. Wavelets have more advantages over compressing signals. Singh and Tomar (2013) states that the wavelet transform is considered as the most advantageous and useful computational tools for a multiplicity of signal and image processing applications. Wavelet transforms are mainly used for images to reduce unwanted noise and blurring. Karami et al. (2012) reported that Wavelet transform has emerged as most powerful tool for both data and image compression. Wavelet transform performs multi resolution image analysis. The DWT has successfully been used in many image processing applications including noise reduction, edge detection, and compression. Indeed, the DWT is an efficient decomposition of signals into lower resolution and details. From the deterministic image processing point of view, DWT may be viewed as successive low-pass and high-pass filtering of the discrete time- domain signal. In 2D image, the images are generally considered to be matrices with N rows and M columns. In wavelet transform, the decomposition of a particular image consists of two parts, the first one is lower frequency or approximation of an image (scaling function) and the second one is higher frequency or detailed part of an image (wavelet function). Figure 2 explains Wavelet Filter decomposition of an image where four different sub-images are obtained; the approximation (LL), the vertical detail (LH), the horizontal detail (HL) and the diagonal detail (HH). Then all the coefficients are discarding, except the LL coefficients that are transformed into the second level.
Back Propagation Neural Networks (BPNN):
The back-propagation learning algorithm is one of the most significant improvements in neural networks. This learning algorithm can be mainly applied to feed-forward networks that can be consists of dispensation elements with uninterrupted differentiable activation functions. In BPNN, training input-output pair is given as an input for training, this algorithm make available a procedure for varying the weights in a BPNN to categorize the given input patterns appropriately. The very vital concept for this weight update algorithm is gradientdescent method. The back-propagation algorithm is completely different from other networks in respect to the process by which the weights are calculated during the learning period of the network. There will be three different layers, one input layer, one output layer and one hidden layer, are assigned. Both of input layer in BPNN and output layer are fully connected to hidden layer. Compression is obtained by designing the value of the number of neurons in the both input layer and output layers neuron less than the hidden layer neuron.
Initially the image is decomposed into numerous pixels using image compression. These pixels are then encoded and given as the input training pattern to the network which is to be transmitted and then reconstructed at the receiver side. In the back propagation process, the entire network consists of input layer, output layer and one or more hidden layers. The spatial co-ordinates of the pixel value are encoded and converted from two to one dimensional value and compressed. When the input is given, it gets multiplied with their corresponding weight to get the total sum of the input. This result of weighted sum undergoes sigmoidal function to yield output pattern. This is the first phase or forward phase. Once the output is gained then the error is calculated after which the process is propagated reversely by finding the changes occur between the output and hidden layer, hidden layer and input layer. Yunong Zhang et al. (2013) propose and investigate two simple structure BP-type neural networks (i.e., the BPNN-L and BPNN-R models) for online generalized matrix inversion. Moreover, two discrete-time Hopfield-type neural networks (i.e., the HNN-L and HNN-R models) are presented and exploited for online solution of the generalized inverse. Panda et al. (2012) proposes this method very clearly and depicted CR, PSNR for Cameraman Image using this BPNN algorithm. In his paper, the compressed image occurs only at 1100, 1300, 1900 epochs and so on.
But our contribution in this paper is that we modified the performance of CR, PSNR from existing paper (Panda et al. 2012). In this paper we got the compressed Image at very less time i.e. within the average of 22 epochs for input medical images. Modification of the performance compared to existing papers can be seen in table 2. The comparison chart is also provided in Fig 11.
Radial Basis Function Neural Networks (RBFNN):
Radial basis function neural networks (RBFNN) are feed-forward networks trained using a supervised training algorithm. They are normally put together with a single hidden layer of units whose output function is selected from a class of functions called basis functions. The structure of an RBF networks in its most basic form involves three entirely different layers as shown in Fig.4. The input layer is made up of source nodes (sensory units) whose number is equal to the dimension N of the input vector. The second layer is the hidden layer which is composed of nonlinear units that are connected directly to all of the nodes in the input layer. Each hidden unit takes its input from all the nodes at the components at the input layer. The hidden unit contains a basis function, which has the parameters center and width.
The general application of this algorithm includes:
* Function approximation in time series modelling
* Classification
* System identification
* Face and Speech recognition
* Image restoration
* Shape from shading
* 3-D object modelling
* Motion estimation and moving object segmentation.
Alex Alexandridis et al. (2013) propose in his paper as, optimization of the Non-Symmetric Fuzzy Partitioning (NSFM) of the input space using PSO (Particle Swarm Optimization) can form the basis for an integrated RBF network training methodology. The algorithm starts by randomly dividing the available data in three datasets, namely, the training, validation, and testing datasets. The existence of a third independent dataset for testing the produced RBF network is crucial, since the optimization procedure could result in over-fitting the model to the validation dataset. During the particle initialization stage, different partitions of the input space are selected randomly, and coded as particles. Each particle then gives birth to an RBF network by applying the NSFM algorithm.
In this paper, the performance of all parameters are modified and got the better Compression Ratio and Peak Signal to Noise Ratio. Thus, compared to all the above algorithms it's depicted that this RBFNN is better Image Compression Technique.
Related Work:
Xingsong Hou et al. (2013) clearly states that the Image Compression using algorithms increase the PSNR up to 3.34 dB and decrease the MPE up to 50.43% as compared with the set partitioning in hierarchical trees (SPIHT) algorithm. DWT algorithm gives better Compression ratio and PSNR to get the good quality of input images. Chun-Lung Hsu et al. (2011) proves that DWT is usually computed through convolution and sub sampling with a couple of filters to produce an approximation low-pass filter result and a detail signal high-pass filter result. The multi resolution decomposition is obtained by iterating the convolution and sub sampling of these two filters over the approximation components. For two-dimension (2D) signals, there exist separable wavelets for which the computation can decompose into horizontal processing (on the rows) followed by vertical processing (on the columns), using the same one-dimension (1D) filter.
Ulug Bayazit (2011) states that a quad tree in DWT represents the spatial partitioning of the set of high frequency coefficients of the colour planes into spatially oriented subsets which may be further partitioned into smaller directionally oriented subsets. The partitioning decisions and decisions to employ fixed or signaldependent bases for each subset are rate-distortion (R-D) optimized by employing a known analytical R-D model for these coefficient coding schemes.
Shiqiang Yan and Zhong Xiao (2013) clearly explain in his paper that there are three layers in the network: input layer, hidden layer and output layer. The network can contain more than one hidden layer in the application. The input-output relationship of the input layer neuron which only receives input is linear. The activation function of output layer neuron is sigmoid function. The hidden layer neuron of this model is different From the neuron of common feed forward neural network. There are two layers in the hidden layer: the first one is composed of F neuron which receives forward input and B neuron which receives feedback input; the hidden layer output really need is the output of the second layer neuron which is called H neuron.
Nicolaos B. Karayiannis and Mary M. Randolph-Gips (2003) states that Training of a cosine radial basis function neural network by gradient descent requires that the active regions of the available radial basis functions cover completely the input space. The definition of the active region indicates that the likelihood of complete coverage of the input space by active regions of radial basis functions improves considerably as the values. The output weights connecting the radial basis functions with the output units were updated according to a supervised procedure based on gradient descent.
Omar Arif and Patricio Antonio Vela (2009) propose in his paper that exploits the relation between kernel methods, regularization theory and radial basis functions to propose a novel method to reduce the computational complexity associated with kernel methods. The work differs from previous efforts in that we approximate the learned function in a manner more compatible with the universal approximation capabilities of neural networks. The final algorithm is a two-step process beginning with the learning procedure, followed by the compression procedure. Efficient representations are achieved with minimal loss of performance.
Srinivasan et al. (2013) states that, Lossless and near-lossless compression algorithms for multichannel electroencephalogram (EEG) signals are presented based on image and volumetric coding. The compression algorithms are designed for principle of lossy plus residual coding. The used methodologies here are Arithmetic coding, compression, electroencephalogram (EEG), multichannel EEG, set partitioning coding. The proposed multichannel compression algorithms achieve not enough compression ratios. he EEG data are first coded at an optimal rate using a wavelet-based scheme, and next the residuals are further encoded by an entropy encoding scheme.
Results:
The quality of compressed image can be measured by many parameters. The most commonly used parameters are Mean Square error (MSE), peak signal to noise ratio error (PSNR), and compression ratio (CR). The PSNR value used to measure the difference between a decoded image and its original image. In general, the larger the PSNR value, the better will be the decoded image quality.
A. Mean Square Error (MSE)
Mean square error is a criterion for an estimator the choice is the one that minimizes the sum of squared errors due to bias and due to variance.
MSE = [1/MN] [[summation].sup.M.sub.X=1] [[summation].sup.N.sub.Y=1] [[f(x,y) - f'(x,y)].sup.2]
Where, f(x, y) is the original image, f(x, y) is compressed image and M, N are the dimensions of the images.
B. Peak Signal to Noise Ratio (PSNR)
It is ratio between sizes of the input image to the square of Mean Square Error (MSE). If PSNR is high then the quality of compressed image is also increased.
PSNR = 10 [log.sub.10] [MXN/[MSE.sup.2]]
Where, M x N is the size of an input image.
C. Compression Ratio
Compression ratio is defined as the ratio between the uncompressed image size and compressed image size. The amount of nose and redundancy removed in an input image can be predicted using CR.
CR = UNCOMPRESSED IMAGE SIZE/COMPRESSED IMAGE SIZE
D. Bits Per Pixel (BPP)
Bits per pixel ratio (BPP) give the number of bits store one pixel of the image.
BPP = SIZE OF COMPRESSED FILE/TOTAL NO.OF PIXEL IN THE IMAGE
The simulation results for images for all the three algorithms are shown below. The images with better compression performance with the better CR and high PSNR are included in the following figures.
The above figures (Fig. 5 to 8) show that the original and compressed images by using DWT, Back propagation Neural Networks (BPNN) and Radial Basis Function Neural Networks (RBFNN) respectively. All the input medical images given here are 512*512 image size. The parameters of the compressed image are depicted using different formulae. By calculating CR and PSNR values of the compressed image we can able to predict the quality of compressed image. Here, by comparing three different algorithms we can conclude that the Radial Basis Function Neural Networks (RBFNN) has better CR and PSNR.
Discussion:
Table 1: Different Parameters depicted for Discrete
Wavelet Transform (DWT).
Algorithm DWT
Parameters CR PSNR MSE BPP Original Image
Size(KB)
Brain Image 1.85 29.28 0.26 1.49 58.5
CT Image 2.76 24.22 0.23 1.58 71.7
MRI Image 3.86 31.96 1.73 1.67 84.0
PET Image 4.73 25.99 0.03 1.89 46.9
Algorithm DWT
Parameters Compressed Image
Size(KB)
Brain Image 19.3
CT Image 21.2
MRI Image 40.3
PET Image 19.9
Table 2: Different Parameters depicted for BPNN algorithm.
Algorithm BPNN
Parameters CR PSNR MSE BPP Original Image
Size(KB)
Brain Image 0.95 49.86 1.47 6.95 58.5
CT Image 0.85 52.46 2.69 8.18 71.7
MRI Image 0.87 49.28 0.58 3.54 84.0
PET Image 0.86 42.52 0.96 7.76 46.9
Algorithm BPNN
Parameters Compressed Image
Size(KB)
Brain Image 17.5
CT Image 21.0
MRI Image 22.0
PET Image 13.0
Table 3: Different Parameters depicted for RBFNN Algorithm.
Algorithm RBFNN
Parameters CR PSNR MSE BPP Original Image
Size(KB)
Brain Image 0.64 60.27 0.93 7.50 58.5
CT Image 0.73 54.29 0.24 7.53 71.7
MRI Image 0.78 58.72 1.65 8.54 84.0
PET Image 0.64 52.56 1.09 10.63 46.9
Algorithm RBFNN
Parameters Compressed Image
Size(KB)
Brain Image 17.0
CT Image 20.7
MRI Image 21.1
PET Image 12.5
Table (1 to 3) gives the parameters values such as CR, PSNR, MSE and BPP of the compressed image for the algorithms DWT, Back propagation Neural Networks (BPNN) and Radial Basis Function Neural Networks (RBFNN) respectively.
The graph is plotted for the values of CR and PSNR for DWT, Back propagation Neural Networks (BPNN) and Radial Basis Function Neural Networks (RBFNN). The final tabulations are also included here after calculating all the parameters used for image compression.
The above figure (Fig9) shows the comparison of Compression Ratio for different input images. This shows that Compression Ratio for DWT algorithm is very high compared to other two techniques. Thus from the graph we can conclude that Radial Basis Function Neural Networks (RBFNN) gives better Compression Ratio.
The above figure (Fig10) shows the comparison of PSNR for different input images. This shows that PSNR for DWT algorithm is very low compared to other two techniques. Thus from the graph we can conclude that Radial Basis Function Neural Networks (RBFNN) gives better PSNR values.
The above figure (Fig11) gives the Comparison of CR Values with Existing BPNN and Proposed BPNN and it's clear from the graph that proposed BPNN has better Compression Ratio. But RBFNN that is also included in this paper gives the better result than proposed BPNN.
Conclusion:
The above presented a comparison of compression techniques based on Back Propagation Neural Network (BPNN), DWT and Radial Basis Function Neural Network (RBFNN) for scan testing to reduce test data volume and test application time. Compressed images for different input images are obtained by these three algorithms. Parameters calculation is also done for the compressed image and comparison chart is provided for Compression Ratio (CR) and Peak Signal to Noise Ratio (PSNR). By comparing the performance parameters of compressed image for these three algorithms, we conclude that RBFNN is efficient technique for image compression. In this paper we review and discuss about the Key issues of image compression, need for the compression, and methods of compression. We review and discuss the advantages and disadvantages of these algorithms for compressing images for Image Compression Techniques. As on future work image compression techniques can be done by combining two or more algorithms to provide hybrid compression techniques.
ARTICLE INFO
Article history:
Received 3 September 2014
Received in revised form 30 October 2014
Accepted 4 November 2014
REFERENCES
Alex Alexandridis, Eva Chondrodima and Haralambos Sarimveis, 2013. "Radial Basis Function Network Training Using a Nonsymmetric Partition of the Input Space and Particle Swarm Optimization." IEEE Transactions on Neural Networks and Learning Systems, 24(2): 219-230.
Chun-Lung Hsu, Yu-Sheng Huang, Ming-Da Chang and Hung-Yen Huang, 2011. "Design of an Error-Tolerance Scheme for Discrete Wavelet Transform in JPEG 2000 Encoder". IEEE Transactions on Computers, 60(5): 628-638.
Karami, A., M. Yazdi, G. Mercier, 2012. "Compression of Hyper spectral Images Using Discrete Wavelet Transform and Tucker Decomposition." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 5(2): 444-450.
Nicolaos, B. Karayiannis and M. Mary Randolph-Gips, 2003. "On the Construction and Training of Reformulated Radial Basis Function Neural Networks" IEEE Transactions on Neural Networks, 14(4): 835-846.
Omar Arif and Patricio Antonio Vela, 2009. "Kernel Map Compression using Generalized Radial Basis Functions." International Conference on Computer Vision (ICCV), 19(9): 1119- 1124.
Panda, SS., MSRS. Prasad, MNM. Prasad and SKVR. Naidu, 2012." Image Compression Using Back Propagation Neural Network." International Journal of Engineering Science & Advanced Technology, 2(1): 7478.
Shiqiang Yan and Zhong Xiao, 2013. "Application of BP Neural Network with Chebyshev Mapping in Image Compression" Third International Conference on Instrumentation, Measurement, Computer, Communication and Control, 7(13): 398-402.
Singh, H.K., S.K. Tomar and P. Singh, 2013. "Analysis of Multispectral Image Using Discrete Wavelet Transform," Third International Conference on Advanced Computing and Communication Technologies (ACCT), 6(7): 59-62.
Srinivasan, K., Justin Dauwels and M. Ramasubba Reddy, 2013. "Multichannel EEG Compression: Wavelet-Based Image and Volumetric Coding Approach", IEEE Journal of Biomedical and Health Informatics, 17(1): 5263-5270.
Ulug Bayazit, 2011. "Adaptive Spectral Transform for Wavelet-Based Colour Image Compression," IEEE Transactions on Circuits and Systems for Video Technology, 21(7): 983-992.
Xingsong Hou, Jing Yang, Guifeng Jiang and Xueming Qian, 2013. "Complex SAR Image Compression Based on Directional Lifting Wavelet Transform With High Clustering Capability," IEEE Transactions on Geosciences and Remote Sensing, 51(1): 527-538.
Yunong Zhang, Dongsheng Guo and Zhan Li, 2013. "Common Nature of Learning between BackPropagation and Hopfield-Type Neural Networks for Generalized Matrix Inversion with Simplified Models," IEEE Transactions on Neural Networks and Learning Systems, 24(4): 579-592.
(1) R. Praisline Jasmi, (2) Mr. B. Perumal, (3) Dr. M. Pallikonda Rajasekaran
(1) PG Scholar, Department of Electronics and Communication Engineering. Kalasalingam University, Virudhunagar, Tamil Nadu, India
(2) Assistant Professor, Department of Electronics and Communication Engineering. Kalasalingam University, Virudhunagar, Tamil Nadu, India
(3) Professor and head of the Department of Electronics and Communication Engineering. Kalasalingam University, Virudhunagar, Tamil Nadu, India
Corresponding Author: R. Praisline Jasmi, PG Scholar, Department of Electronics and Communicationn Engineering, Kalasalingam University, Virudhunagar, India. Address: 2/602, Maranatha Church, Bharathi Nagar, Reserve Line, Sivakasi-West. Pin Code:626 124.
Tel: +91 96004 56608 E-mail: praise.jas@gmail.com
Fig. 2: Three Level Decomposition Wavelet Filter.
LL3 LH3
HL3 HH3 LH2
HL2 HH2 LH1
HL1 HH1
 Printer friendly
 Cite/link
 Email
 Feedback
| |
| Title Annotation: | discrete wavelet transform |
|---|---|
| Author: | Jasmi, R. Praisline; Perumal, B.; Rajasekaran, M. Pallikonda |
| Publication: | Advances in Natural and Applied Sciences |
| Article Type: | Report |
| Geographic Code: | 9INDI |
| Date: | Nov 1, 2014 |
| Words: | 4038 |
| Previous Article: | A novel approach for detecting rare drug-drug interaction using data mining. |
| Next Article: | Design of multi output binary adder using modified parallel prefix addition. |
| Topics: | |