IMAGE PROCESSING USING NEURAL NETWORKS

Abstract

Artificial neural networks (ANNs) are very general function approximators, which can be trained based on a set of examples. Given their general nature, ANNs would seem useful tools for nonlinear image processing. This paper explores the application of neural networks in digital image processing. Particularly, it details the design and implementation of a neural network based edge recognizer, and studies the effects of incorporating prior knowledge in its design. After a brief introduction to ANNs, ANN training algorithms, and digital image processing, the paper explains the design of a neural network based edge-recognizer and its training using the conjugate gradient descent (CGD) training algorithm. When the network is trained using the aforementioned training algorithm, it was observed that ANNs can be used as edge detectors. However, the presence of receptive fields in the architecture in itself does not guarantee that shift-invariant feature detectors will be found. The appendix at the end of the report contains the implementation source code. 8