PCNN And One-Class SVM For Texture Retrieval

Content-based image retrieval (CBIR) has been researched and used extensively in image retrieval. The key technology of CBIR is translation, rotation and scale invariant feature extraction. So image retrieval based on invariant feature has broad research prospects. The outputs of pulse-coupled neural networks (PCNN) contain unique characteristics of the original stimulus, which is especially suitable for invariant feature extraction. Support vector machines (SVM) solved problems like the limited samples and the curse of dimensionality, which can obtain global optimal solutions. Especially the succeeding one-class SVM get a good solution to the one class problem and are widely used in image retrieval. In this paper, for studying the extensive research on the existing texture image retrieval, the algorithm of training pulse features based on one-class SVM is illustrated. The main studies of this paper are as follows:(1) The contents, methods and development status of the invariant feature retrieval are introduced systematically. The key technologies of invariant feature retrieval are analyzed. The basic principles and applications of PCNN and one-class SVM are learned and explored. The above works provide a theoretical basis for the algorithm of this paper.(2) The geometric invariant texture retrieval system based on PCNN and one-class SVM is proposed. First, PCNN is used to extract invariant features on the popular texture images, and then training those pulse features of PCNN based on one-class SVM. In experiments taking different texture images with rotation and scaling changes for retrieval, and the experimental results show that the proposed retrieval system has a good rotation and scale invariance.(3) Taking anti-noise experiments of texture retrieval system based on PCNN and one-class SVM. The anti-noise experiments are done by adding different noise to the texture images, and the experimental results show that the proposed retrieval system is quite robust to noise.It arrives at a conclusion that the combination of pulse features and statistical learning theory has a broad application prospects in image retrieval.