Rule Insertion and Extraction in Evolutionary Neural Networks for Image Retrieval

点击翻译按钮获取中文摘要

This study presents a new knowledge incorporation and rule extraction method to deal with knowledge in neural networks for image retrieval in the Internet. The rule form of an if-then type can be inserted into a neural network (NN) as knowledge of a problem. NN is then trained by using a set of training samples. In this case the structure learning algorithm with forgetting is used to generate a small-sized NN system. After the NN training, rules are extracted from it. Furthemore evolutionary methods are used to train the neural network structure. The results of computer simulations for pattern recognition and chaos show that this approach can generate obvious network architectures and as a result simple rules compared with conventional rule extraction methods.On the one hand, a method for image classification by neural networks which uses characteristic data extracted from images is studied. And also rule extraction from images has be done by neural network learning. Accuracy for image classification and key word extraction is about 70%.In the future, these techniques must be unified and implemented for image retrieval.