Structured and random texture patterns characterization using multiscale directional filter bank

The use of multiscale directional decomposition, achieved by combining a Laplacian pyramid and a directional filter bank, is studied for texture classification. We first demonstrate the importance of the multiscale analysis of directional texture features. Then, it is found that directional analysis is suitable for characterizing structured textures, but not random textures. Thus, structured and random textures are separated by employing an entropy-based measure on the multiscale directional features. Through this pre-filtering step, structured textures are extracted for further classification, so that the overall retrieval performance can be enhanced. Experimental results showed that this pre-filtering step can significantly improve the overall retrieval accuracy.