Multiple channel neural network model for texture analysis

With the recent progress in artificial neural network (ANN) research, the performance of these networks in image processing and analysis operations is interesting and important. A major issue in applying these networks to image problems is the representation of the data presented to the network. For manageable training and computation, data reduction and feature representation is essential. A modular image analysis model is proposed which incorporates ANN and more conventional image processing techniques, with particular attention being paid to the representation of data and the utilization of output signals from ANN.; The multiple channel neural network model is comprised of four basic modules: image preprocessing normalization, multiple channel representation filters, a neural network classifier and a contextual correction scheme. To carry out multiple channel representation, Gabor filters are employed. To generate tuned Gabor filters for better image representation, a two-phase filter adaptation algorithm is introduced.; Various problems in textural image analysis are chosen to evaluate the performance of the model. In fingerprint features identification, the system is able to locate minutia positions in fingerprint images. In textural image classification, both scale-rotation variant and invariant classifications are carried out. Simulations demonstrate the usefulness of neural networks and Gabor filters as classifiers and image features as the model gives better performance when compared to other approaches. In textural image segmentation, a context competition algorithm is introduced to utilize image context to improve segmentation results of composite textural images.