Texture Detection Using Neural Networks Trained on Examples of One Class

摘要

We describe an approach to finding regions of a texture of interest in arbitrary images. Our texture detectors are trained only on positive examples and are implemented as autoassociative neural networks trained by backward error propagation. If a detector for texture T can reproduce an n x n window of an image with a small enough error then the window is classified as T. We have tested our detectors on a range of classification and segmentation problems using 12 textures selected from the Brodatz album. Some of the detectors are very accurate, a small number are poor. The segmentations are competitive with those using classifiers trained with both positive and negative examples. We conclude that the method could be used for finding some textured regions in arbitrary images.