Research On Image Recognition And Its Application Based On Deep Belief Network

Deep Belief Network is a new model of machine learning in recent years. Its motivation is simulating human intellection to learn and analyze data, such as text,sounds and images. Due to using a combination of unsupervised pre-training and supervised fine-tuning, DBN has the advantage of automatically extracting samples' probability distribution and capturing samples' essential features, so as to realize image recognition with big data.In this thesis, depending on available theory of DBN, it made a deeper study of improved recognition algorithms and its application based on DBN. The contents of this paper mainly focus on the following aspects.Based on DBN, two improved recognition algorithms were put forward in view of the defects of the recognition performance is not too high. One improved DBN recognition method is based on multi-scale main direction features. Firstly, the main features were extracted from sample images according to the extracting flow of multi-scale main direction features. Then we inputted amplitude information into the network of DBN, along with the main features which were used as the original amplitude information's instructions. Another improved DBN recognition method is based on sparse difference. At first, the concept of difference was defined. After that, the gray values of image pixels were converted into difference values, and achieved the goal of expanding low gray domain, compressing high gray domain, enhancing the contrast of images. Based on this, the removing mean value normalization and sparseness is processed to the difference data. Then, the difference matrixes were inputted into DBN.The recognition results on MNIST, CIFAR-10 and SVHN database by comparing with other algorithms show that, both the two improved methods can improve the recognition performance of DBN efficiently.Then, through the study and application of fault indicators inspection and the fault identification for insulators in enterprise production, a significant amount of experiments were carried out with improved image recognition method of DBN. The results show that the two practical applications acquired satisfied effects in recognition. Furthermore, the effectiveness and practicality of the improved method were proved. Finally, an improved image segmentation algorithm based on the image sharpness theory was proposed,accordingly realized images' preprocessing before DBN recognition.