Image restoration based on Hopfield neural network and wavelet domain HMT model

The traditional image restoration algorithms based on a Hopfield neural network are unable to compress the noise and protect the details at the same time. In order to solve the problem, a new algorithm based on the modified Hopfield neural network with a continuous state change and the wavelet domain Hidden Markov Tree (HMT) model is presented. The wavelet domain HMT model is utilized as the prior information about the statistical relationship between the two image wavelet coefficients, and is introduced into the neural network model by a regularization term. The final restoration image is obtained by using the energy convergence property of the Hopfield neural network. Furthermore, a highly-parallel weight matrix determination algorithm is proposed, and then the weight values are computed batch by batch through the operation to the pattern images to avoid the multiplication of large scale matrices. Experimental results demonstrate that the visual quality of the restoration result is improved evidently for either real images or artificial images, and the Improved Signal to Noise Ratio (ISNR) is improved more than 0.3 dB compared to that of the traditional algorithms. The objectives of compressing the noise and protecting the details are achieved at the same time.