Image deblurring based on wavelet and neural network

A new method of image deblurring based on wavelet decomposition and RBF (Radial Basis Functions) neural network is proposed. Traditional algorithms of image deburring have some limits, such as assumption of generalized stationary process and calculating complexity etc in their application. Based on the data of degraded image and original image as trained samples of the RBF network, the similarly degraded image may be restored through recording the reverse process. Combining wavelet transform with neural network, through wavelet decomposition of degraded image and original image, images of different low frequency resolution may be obtained; thus, multi-resolution image pyramids of the two sample images can be constructed. The RBF network structure may be optimized by levels through the sample images trained from the top to the bottom of the pyramid, and the weights obtained from the upper level were used as the original weights of the next level. The image restoration thus can be conducted by the eventually obtained optimized network. Simulation results show that the proposed method is easily achieved, the convergence is accelerated and the ideal deblurred result is realized.