The Research Of Image Classification Methods Based On Convolution Neural Network

Convolution neural network which is a kind of pattern recognition methods combining artificial neural network and deep learning theory arises in recent years and has become one of the focus areas of the image classification. Different from traditional image classification algorithms, convolution neural networks don’t require the extraction of specific image features for specific tasks, but simulate the human visual system to provide a hierarchical abstraction for original image to generate classification results. The method make use of local receptive fields, shared weights and spatial sampling technique, making significantly reduced training parameters of the network compared with traditional neural network and bringing a certain degree invariance against translation, rotation and distortion, and has been widely applied in speech recognition, face recognition, handwriting recognition, pedestrian detection and other applications. Compared with traditional image classification methods, the application of convolution neural network in image classification field will bring higher recognition rate and wider applicability. Therefore, the study of convolution neural network has important theoretical significance and applying convolution neural network to handwritten digit recognition, face recognition, identification of plant leaves and other research fields will also bring important practical application significance.In this thesis, we begin with the basic concepts and algorithms of convolution neural network and make a in-depth study of convolution neural network theory, aiming at improving its fixed structure to propose a new algorithm and improving its performance in handwritten digit recognition, face recognition and plant leaf identification applications.(1) We summarize the latest research on artificial neural networks and convolution neural network, introduce the basic concepts of convolution neural network and basic principles. Furthermore, we expatiate its basic structure and network parameters, point out some advantages and disadvantages of convolution neural network.(2) Since conventional convolution neural network architecture is used in the handwritten digit recognition, we make some fine-tuning on the network structure to accommodate the number of categories as well as the size of the face image of a specific face database, and design the convolution kernel and excitation function to adapt to the system. The experimental results show the identify performance of convolution neural network is more efficient than the traditional pattern recognition methods.(3) Regard to problems that the conventional convolution neural network lack of a network configuration design theory for specific issues, network design is often based on experience according to the specific problem, and its structures are set by artificial, its number of neurons and connections are fixed cannot make any adjustment. We presents a dynamic growth convolution neural network architecture to automatically find a reasonable structure suited to the application requirements. In addition, regard to the over-fitting problem existed in the convolution neural network that excessive training but to make the identification of performance degradation, we use a active sample learning method to construct an effective training set. The combination of these two methods is verified by experiment that it can achieve better results than conventional convolution neural network.