Generalized Probabilistic Decision-Based Neural Networks for texture classification and retrieval

For various applications, formulating texture features in distributional forms can sometimes provide meaningful representation than in numerical forms. In this paper, we first proposed a novel methodology for measuring the difference between two mixture Gaussian distributions. Based on the derived formula, a Generalized Probabilistic Decision-Based Neural Network (GPDNN) is then proposed and implemented to realize the difference measurement method. By constructing a two layered pyramid-type network structure, the proposed GPDNN receives data in distributional form via 2-D grid input nodes, and outputs the classification and/or retrieval results from the top layer node. Forty texture images are selected from the MIT Vision Texture (VisTex) database to evaluate the proposed GPDNN for the texture classification and retrieval. Experimental results show that (1) by using the proposed difference measurement methods, the texture pattern retrieval rates can be improved from 77% to 82%, compared with some published leading methods, and (2) the proposed GPDNN shows significant texture classification and retrieval performance, which are about 90:1% and 88:6% of the accuracy, and much better than the traditional methods, i.e., 82.2% and 79.9%, respectively.