In this paper we present a technique for localized image regularization using a Modified Hopfield Neural Network (MHNN). The algorithm forms a segmented map of the image and classifies it into several clusters, or regions, and assigns each region a regularization parameter according to its local statistics and the prior knowledge about the image obtained by a Bayesian Minimum Risk (BMR) restoration method. The image segmentation is performed over the BMR restored image. First, the user selects arbitrarily at least one region, and makes a subjective decision to choose the best estimate from among a set of restored images with different regularization parameter applied to the user-selected region. Then, using this decision the algorithm sets up a perception-based selection of the different regularization parameters for restoring in an adaptive fashion the whole image employing the MHNN computations.
Spatially Adaptive Regularization Image Restoration Using a Modified Hopfield Network
2007-09-01
691404 byte
Conference paper
Electronic Resource
English
Adaptive regularization in image restoration by unsupervised learning
British Library Online Contents | 1998
|An Adaptive Segmentation-Based Regularization Term for Image Restoration
British Library Conference Proceedings | 2005
|Optimization of regularization operators for adaptive least squares image restoration
British Library Conference Proceedings | 1994
|A Hopfield neural network for adaptive control
AIAA | 1993
|