Adaptive regularization in image restoration using a model-based neural network

Abstract

Regularization parameter estimation is an important issue in the overall optimization of image restoration systems. The parameter controls the relative weightings of the data- and model-conformance terms in the restoration cost function. In general, we should adopt small parameter values for highly textured image regions to emphasize detail and should use large values to suppress noise in smooth regions. In spite of this qualitative knowledge, the exact value of the parameter is normally difficult to estimate due to the nonintuitiveness of the parameter value as an indicator of the resulting image quality. In view of this problem, we propose a regionally adaptive regularization approach that first specifies the desired image regional quality in terms of a specific predictive filter mask and then establishes a correspondence between the filter mask and a regional regularization parameter value using a modelbased neural network. Due to the ease of tailoring local image quality using a filter mask by judiciously specifying the mask coefficients, we can define separate filter masks for different image regions to indicate different preferences for detail preservation. We can then relate the prediction given by each filter mask to a specific parameter value through the function approximation capability of the model-based neural network and fine-tune this value through training. As a result, the current assignment is more relevant in relation to the local spatial characteristics of the image than with the usual practice of using an arbitrary function of the SNR to determine the parameter value.