Abstract
The explosive growths in data exchanges have necessitated the development of new methods of image compression including use of learning based techniques. The learning based systems aids proper compression and retrieval of the image segments. Learning systems like. Artificial Neural Networks (ANN) have established their efficiency and reliability in achieving image compression. In this work, two approaches to use ANNs in Feed Forward (FF) form and another based on Self Organizing Feature Map (SOFM) is proposed for digital image compression. The image to be compressed is first decomposed into smaller blocks and passed to FFANN and SOFM networks for generation of codebooks. The compressed images are reconstructed using a composite block formed by a FFANN and a Discrete Cosine Transform (DCT) based compression-decompression system. Mean Square Error (MSE), Compression ratio (CR) and Peak Signal-to-Noise Ratio (PSNR) are used to evaluate the performance of the system.
TopIntroduction
The last few decades have seen unprecedented growth in communication and multi-media technologies. Both of these have been symbiotically driven. The use of multimedia contents in every segment of day to day life has necessitated the urgent requirement of providing more and more storage space. This requirement, to a large extent, can be optimized by use of compression technologies. This has motivated the research in digital image compression domain. Of late, the importance of application of innovative technologies has reached a new level. Though traditional approaches have already established their importance and effectiveness, still there are ample of opportunities to explore new methods of achieving compression. Modern day vision devices have placed great leverages on resolution which have constantly increased the data volume. Concurrently, though compression methods have supported the requirements of more and more storage space and communication capacity, improvement in decompression and presentation approaches have become stringent factors. Therefore, design of improved compression and decompression methods have become a need of the hour. A practical compression algorithm for image data should preserve most of the characteristics of the data while working in a lossy manner and maximize the gain and be of lesser algorithmic complexity (Ettaouil, Ghanou, Moutaouakil, & Lazaar, 2011; Durai & Saro, 2006). The process of image compression is based on reduction of redundancy contained in an image. Some of the benefits derived from image compression are cost savings associated with sending less data over switched telephone or wireless networks, reduction of storage requirements and overall processing/execution time, reduction in transmission errors and improvement in security against illicit monitoring. So while image compression provides several benefits, complex methods are required to restore the forms before presentation. It is more so if medical images are involved.
Medical images are needed to be stored in diagnostic centers for a patient’s future reference and also for the purpose of transmission through a communication channel for telemedicine and referral for advanced treatment. Hence, it has to be compressed before using it for storage or transmission. Learning systems like Artificial Neural Networks (ANN) can be used to pre-process input patterns to yield simpler patterns with smaller components. This motivates the users to prefer it for image compression. ANN based image systems provide high compression rates and preserve quality. They have the ability to preprocess input patterns to produce reduced data sizes with fewer components. Self Organizing Feature Map (SOFM) is a kind of ANN which consists of components called nodes (neurons) and is based upon unsupervised learning. It gives a low dimensional and discretized representation of the input space of the training samples which is known as map (Sharma & Kumari, 2013). The SOFM model proposed by Kohonen can be used for image compression. The primary principle behind the use of SOFM is the fact that it is inherently a data reduction method. Considerable compression techniques have been designed, regarding the use of conventional and ANN based methods for compression of images. Here, we include learning and decomposition based image compression technique using ANN in feedforward (FF) configuration and applied as SOFM. The work describes a lossy compression system for compression of medical images. An image is first decomposed into smaller blocks before passing it through FF ANN and SOFM network for generation of a codebook. The output of FF ANN and SOFM are further compressed by applying Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT) like techniques to it. The compressed image is then reconstructed using a multi-layer FF ANN. The performance of the proposed system is judged by finding parameters like Mean Square Error (MSE), Compression Ratio (CR) and Peak Signal-to-Noise Ratio (PSNR). The original image is degraded to some extent by applying salt and pepper noise to it and applied to the proposed system. The PSNR value of the reconstructed image containing noise is determined. It is found that the system is robust to degradations like noise that may be present in an image due to false switching of the device and transmission through a noisy channel.
Key Terms in this Chapter
Supervised Learning: It is a type of learning where the systems follow a pre-determined pattern. It is like learning with the support of a teacher. It is generally observed in MLP type ANN.
Unsupervised Learning: It is used with SOM where there is no reference to follow. It adopts a competitive learning approach to form clusters of samples that have commonality.
Multi Layer Perceptron (MLP): Is an ANN type that requires a reference to learn patterns. It is trained using (error) back propagation algorithm.
Lossless Compression: It is technique of image compression which has a lower compression ration but provide high decompression quality. It is suitable for medical applications.
Self Organizing Map (SOM): Is a special form of ANN that uses a competitive learning approach to perform classification and clustering. It requires no referencing.
Lossy Compression: It is method of image compression where the compression ratio is high but the quality is poor. It is useful for less critical applications like text, voice etc compression and storage.
Discrete Wavelet Transform (DWT): Is a wavelet transform performed in discrete domain. It has mother and daughter wavelets with which multi level decomposition is achieved and spectral representation of samples obtained. DWT is often used in image compression and restoration applications.
K-Means Clustering (KMC): Is a grouping technique where objects with common characteristics are placed in K-numbers of identical groups considering the sample to sample difference of mean to be a cost functional.
Mean Square Error (MSE): Is a cost functional used to ascertain the level of training that a prediction system achieves. MSE is to measure the closeness between present output to actual output.
Artificial Neural Network (ANN): Is non-parametric tool that learns from the surroundings, retains the learning and uses it subsequently.
Discrete Cosine Transform (DCT): Is a forward transform used in image compression. It separates DCT coefficients into most relevant and least relevant types.