Texture Detection using Neural Networks Trained on Positive-Only Examples

The task of texture classification is to determine the category of a texture image. Most texture classifiers have been trained on a training set of positive and negative examples, with purpose to learn to separate the positives from the negatives. The possible negatives are usually unlimited. The performance and accuracy of texture classifiers therefore largely depend on the choice of negatives. This work investigates training texture classifiers trained only on positives, without negative examples. We propose a methodology to detect texture using auto-associative neural networks. Brodatz textures are used to perform the experiment. Texture images are processed on pixel level, with no features extraction. Created detectors can also be used to segment different images that contain texture regions in simple and complex shapes. This work presents a unique approach to texture classification, and successfully detects single textures in different random environments. For some textures, our detectors are very accurate, while the detectors for other textures are only fair. Our texture classifiers are able to detect 96% of positive examples, and are mostly effective in rejecting other textures. While our training conditions are more limited, our detectors achieved comparable results to Song(2003). We suggest that more accurate classifiers can be found by using larger window size of example data. When being used for segmentation, the classifiers can successfully segment textural regions in different images. The applicability of classifiers in different image types without the need of updating information about surrounding background is the strength of our classifiers.