Image super-resolution reconstruction based on self-similarity and neural networks

A novel super-resolution approach is presented. An image pyramid has been built based on the framework of wavelet transform, and the detailed coefficients are explored for training the neural networks. The initial high resolution image is estimated by the trained networks and the inverse wavelet transform, and then is constrained with prior knowledge of the error function by iteration. For a factor of 2n, repeat this process and update the networks. The experimental results show that our method reconstructs the more reliable image without obvious visual artifacts.