Theoretical foundation for nonlinear edge-preserving regularized learning image restoration

Abstract

Image restoration is an important issue in image processing, which helps recovering of degraded images caused by various factors in different circumstances. Neural networks models have been successfully applied in handling image restoration problems and some progress and promising results have been reported in the past. A popular neural networks model for image restoration is the Hopfield network due to its ability on dealing with optimization problems. A degraded image may have multiple corresponding solutions, i.e., the restored images, and the obtained solution can be sub-optimal which is related to a local minimum point in the weight space of the Hopfield network. This paper gives an algebraic characterization on images and shows that images with lower complexity can be restored uniquely from their degraded images and edge information. The obtained result in this paper establishes a mathematical basis for employing the mapping neural networks to realize a learning based image restoration scheme. The linear image restoration model is firstly generalized to a nonlinear one to broaden the scope of application. Secondly, we view the image restoration task as a set of approximation problems in a high dimensional space, and a mapping relationship between the degraded images with edge information and the source images is then built using feed-forward neural networks and the well trained neural network can be used to restore the degraded images in realtime. Computer simulations demonstrate the effectiveness of the learning image restoration techniques proposed in this paper.