Research And Application On Image Recognition Method Based On Deep Learning

Image recognition is an important research direction in the field of image research,and it is also a hot research problem in machine vision, which is of great significance.In recent years, deep learning has made many achievements in the field of images,voice, text and so on. At the same time, deep learning occupies an important position in the field of artificial intelligence, which is widely applied and concerned in daily life.The traditional method of image recognition requires the feature of artificial design,which relatively depends on the research scholars with rich experience of image recognition. Moreover, the rate of image recognition of the traditional method is low.With the development of Internet and information technology, with the generated massive image data under the background of big data, the traditional identification method has not been able to meet our needs. However,deep learning technology is a multi-layer network structure. Through the simulation of human brain, deep learning algorithm can automatically learn and extract features, which fully takes the advantages of big data. Therefore, this paper combines the deep learning and image recognition to study how to improve the rate of image recognition, which has a considerable research potential and research value.This paper firstly introduces theories of image recognition and deep learning,compared with shallow learning, deep learning can easily express complex functions and has a strong generalization ability. At the same time, we also discuss several commonly models of deep learning and their principles of algorithms, and we study the feature extraction and recognition method of image.Based on the study of deep neural network, in this paper, we propose an improved method of initializing weight on the problem of slow learning speed in network caused by the original method of initializing weight. At the same time, we prove the validity of the method in theory and experiment, and it can be applied to the commonly convolutional neural networks and deep belief networks.Then, due to the problem that is disappeared gradient in deep neural network. At the same time, deep belief networks have a characteristic that is semi-supervised learning and can dig the value of a large number of unlabeled data. So, we propose improved deep belief networks in this paper. Experiments show that the learning speed and the recognition rate of this model are improved. Compared with unimproved deep belief networks, the recognition rate of this model on the MNIST dataset is 99.18%,increased by 0.62%, and the recognition rate on the CIFAR-10 dataset is increased by 9.6%.Finally, because convolutional neural networks are particularly suitable for dealing with problems that are related to images, we propose improved convolutional neural networks in this paper. This model firstly replaces the originally initializing method with the improved method of initializing weight; then we removes the pooling layer and replaces the original softmax layer with SVM classifier; lastly we improves the activation function which combines the smoothness of the Sigmoid function and the sparseness and fast convergence of the ReLU function, at the same time, we introduce the idea of Dropout to enhance the generalization ability of the network and prevent the network from over-fitting. The recognition rate of this model on the MNIST dataset is 99.52%, which is 0.66% higher than unimproved convolution neural network,which is about 5% higher than traditional methods. In the CIFAR-10 dataset,the recognition rate is improved by 6.4% compared with unimproved convolution neural network, which is about 9% higher than traditional methods. Experiments show that the effectiveness of this model is verified, this model show better performance and improve the rate of image recognition.