Unified Entropy in Self-organizing Feature Maps Neural Network

摘要

Pattern Recognition is a very urgent research area in intelligent information processing and computer intelligent perception, such as computer vision, content-based retrieval, image-processing, etc. In general, the research on pattern recognition is carried out partial separately as feature extraction, classification, etc. in which samples of feature extraction could not be reliable and the global optimum could not been achieved. In this paper the unified entropy theory on Pattern Recognition is presented firstly, in which the information procedures in learning and recognition and the determine role of Mutual Information have been discovered. Secondly build SOFM neural network and apply Mutual Information entropy to compute reliability of training samples, through which selecting excellent data samples is presented to get optimum recognition performance, which is crucial for difficult pattern recognition problems. Experiments on device state recognition prove their effective and efficient.