Image compression using neural network

Abstract

In this work, neural network architecture is used for image data compression. It is well suited to the problem of image compression due to their massive parallel and distributed nature. As the digital images require remarkable amount of memory capacity on the disk and also large amount of bandwidth for transmission; therefore image compression algorithms play a significant role in real life applications. One of the successful applications out neural network applications is Principal Component Analysis (PCA) which is used for the image data reduction. PCA is a mathematical technique to transform input data set into lower dimensional space and retain most fundamental data of the original image. In this work, we analyze PCA method with the help of neural network in which free parameters; the synaptic weights act as the principal components which are trained through the iterative method technique known as Generalized Hebbian Algorithm (GHA). A Comparison with the traditional PCA methods is also performed to demonstrate and illustrate the training and capabilities of Generalized Hebbian Algorithm for image data compression. Simulated data is also presented to evaluate their performance. The evaluated results show that GHA method with neural network architecture gives promising results as the number of iterations increases over both the traditional PCA methods; namely, Singular Value Decomposition (SVD) and Gram Schmidt Orthogonalization Procedure (GSP), in terms of Mean Square Error (MSE), Peak Signal to Noise Ratio (PSNR) and Structure Similarity Index (SSIM).