ADAPTIVE REGULARIZATION IN IMAGE-RESTORATION USING A MODEL-BASED NEURAL-NETWORK

Regularization parameter estimation is an important issue in the overa ll optimization of image restoration systems. The parameter controls t he 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 shou ld use large values to suppress noise in smooth regions. In spite of t his qualitative knowledge, the exact value of the parameter is normall y difficult to estimate due to the nonintuitiveness of the parameter v alue as an indicator of the resulting image quality. In view of this p roblem, we propose a regionally adaptive regularization approach that first specifies the desired image regional quality in terms of a speci fic predictive filter mask and then establishes a correspondence betwe en the filter mask and a regional regularization parameter value using a model-based neural network. Due to the ease of tailoring local imag e quality using a filter mask by judiciously specifying the mask coeff icients, we can define separate filter masks for different image regio ns to indicate different preferences for detail preservation. We can t hen relate the prediction given by each filter mask to a specific para meter 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 loc al spatial characteristics of the image than with the usual practice o f using an arbitrary function of the SNR to determine the parameter va lue. (C) 1997 Society of Photo-Optical Instrumentation Engineers.