Edge and Texture Detection Using Gabor Filters Through Neural Network for Image Classification - Abstract

Abstract

            This work describes detection of edges in images and texture using Gabor filters. In this work edge detection is based on quadtrees and a classification technique in the Gabor transform domain. In this scheme, a quadtree is used to segment low-detail regions into variable sized blocks and high-detail regions into uniform 4*4 blocks. High-detail blocks are classified by an edge-oriented classifier which employs a multilayer neural network with edge models defined in the normalised Gabor domain. The proposed classifier is a perceptron multiplayer Neural Network with Back-propagation learning algorithm.

            The human’s capability to distinguish perceptually different textures is difficult to reproduce using machine vision due the variety of textural patterns and illumination conditions. Here we have proposed a multi-channel texture analysis technique that relies on 2-D Gabor filters to isolate regions of perceptually homogeneous texture in an image.Textures are modeled as a pattern dominated by a narrow band of spatial frequencies and orientations. Properly tuned Gabor filters react strongly to specific textures and weakly to all others. The amplitude of the channel outputs is compared to segment the textures. The phase of the channel outputs is used to locate discontinuities in the texture phase.