Blind image restoration using multilayer backpropagator

Abstract

We describe the problem of restoring a blurred and noisy image without any prior knowledge of the blurring function and the statistics of additive noise. A multilayer feed-forward neural network based on backpropagation algorithm is used for image restoration. The neural network is trained by applying backpropagation with momentum for fast convergence. The results of the backpropagation neural network model are compared to that of Wiener filter for high, moderate and low signal to noise ratio (SNR) blur functions. Improvement in signal to noise ratio (ISNR) is taken as a performance measure. It is observed that backpropagation neural network learns well in each case and restores all the test images reasonably, while Wiener filter performs well for high and moderate SNR blur but performs poorly for the low SNR case. ISNR values of 5.58 db, 5.15 db and 5.13 db has been achieved with this scheme for the peppers image, in comparison to values of 4.17db, 2.71db and -0.93db using Wiener filter for high, moderate and low SNR blur respectively.