Tri-Training Based Deep Learning Network For Image Classification

Over the past decade, the deep learning algorithm is an effective structure for character representation, which have performed better than traditional network in feature extraction, data analysis, image processing and natural language understanding. The training algorithm for deep learning has solved the optimization problems for multi-layers perceptron models, which has a more powerful performance on feature extraction. Among them, the deep convolution neural network with local sensor and shared weights has an advantages in local features understanding and training time. The deep learning networks including convolution network has played an extremely important role in big image data processing tasks.Based on the deep understanding of the basic principles and network structure for a large number of deep learning algorithm, this paper summarize and introduce the network structure and principles for the deep convolution neural network. Furthermore, we apply the deep convolution neural network to the actual image recognition task. In addition, taking into account the actual scene classification task complexity, in this paper, we combine the tri-training algorithm and convolution neural network to construct a common goal of sustained automatic classification method which can learning new features automatically.The main work is as follows.1) Starting with the understanding of the principle of BP neural network algorithm, this paper mainly studied the structural characteristics of the convolution neural network and summarized the training process using gradient descent algorithm.2) We apply the deep convolution neural network to the handwritten image recognition experiment in order to verify the powerful character representation of convolution neural network for image data. Additionally, we observed the effect of different network parameters to the neural network.3) Combining the tri-training algorithm and convolution neural network, we construct a common goal of sustained automatic classification method which can learning new features automatically. We improve the robustness of deep tri-training network by optimizing the mark of confidence and denoising training algorithm.4) We apply the deep tri-training network to Gender recognition and classification for pedestrian and vehicle experiments. By contrast with the traditional tri-training algorithm, we summed up the advantages and disadvantages of our method.