Multistage neural networks and multiresolution approaches to signal and image processing

The Hopfield neural configuration has been employed in a partitioned mode to achieve signal restoration. This approach is computationally efficient and capable of progressing to deeper local minima on the reconstruction error surface. Other features include increased parallelism and ease of hardware implementation. We propose two distinct multistage algorithms to perform error minimization. The partial data approach treats all the neurons of subsets of the data while the partial neuron strategy divides the data set into regions based on the neurons of the entire data. We analyze the performance of these two methods in the context of image restoration. In an effort to understand low level biological vision processes, we have investigated the edge localization features of bandpass masks in the transform domain. The spatial frequency localized Laplacian of Gaussian filter in conjunction with the Discrete Cosine Transform yields superior edge positional accuracy when compared with traditional techniques. This has significant ramifications in several areas including motion compensated image coding and contour tracking in medical images. In the latter case, we have used 2-D and 3-D extensions of the filter to perform object boundary detection in Magnetic Resonance image scans. A systematic selection and variation of the filter parameter controls the accuracy of the contour map and reduces edge misclassification. Further, a parallel implementation of this approach makes it a powerful tool in image data analysis