An Image Classification Method Based on Deep Neural Network with Energy Model

Yang Yang, Jinbao Duan, Haitao Yu, Zhipeng Gao, Xuesong Qiu
2018 CMES - Computer Modeling in Engineering & Sciences  
The development of deep learning has revolutionized image recognition technology. How to design faster and more accurate image classification algorithms has become our research interests. In this paper, we propose a new algorithm called stochastic depth networks with deep energy model (SADIE), and the model improves stochastic depth neural network with deep energy model to provide attributes of images and analysis their characteristics. First, the Bernoulli distribution probability is used to
more » ... lect the current layer of the neural network to prevent gradient dispersion during training. Then in the backpropagation process, the energy function is designed to optimize the target loss function of the neural network. We also explored the possibility of using Adam and SGD combination optimization in deep neural networks. Finally, we use training data to train our network based on deep energy model and testing data to verify the performance of the model. The results we finally obtained in this research include the Classified labels of images. The impacts of our obtained results show that our model has high accuracy and performance. Convolutional neural networks have obtained a large number of applications at present, such as image recognition, image target detection, image segmentation and other fields. Especially in the ImageNet competition after 2012 (this competition is mainly to identify a data set with 1000 categories), basically every contingent team uses the convolutional neural network extensively. At the beginning of the game, the team that used the convolutional structure of the neural network received the absolute leading result of the first place. The error rate of this team was half that of other teams such as the SVM algorithm. Therefore, most image processing tasks today use convolutional neural networks. However, these methods exist problems of insufficient image recognition accuracy and scalability. Although Lv et al. [Lv, Yu, Tian et al. (2014) ] have studied networks using DBN (Deep Belief Network) neural network, the problem of training difficulty still exists in spite of the fact that many works show network depth is very influential [Szegedy, Liu, Jia et al. (2015) ]. Besides, many scholars have carried on continuous research on this problem before [Ioffe and Szegedy (2015) ; Jaderberg, Czarnecki, Osindero et al. (2016) ], as we stack more layers, such as over a hundred layers, the optimization of deep neural networks has been approved to be more difficult, the obstacle problem is vanishing gradients [Bengio, Simard and Frasconi (1994) ; Glorot and Bengio (2010) ; He, Zhang, Ren et al. (2015)], which hamper the increase of neural layers [Saxe, McClelland and Ganguli (2013); Ioffe and Szegedy (2015); He and Sun (2015) ]. Stochastic depth network can successfully address these problems [Huang, Sun, Liu et al. (2016) ], it can greatly increase the nets depth and bring substantial performance than former networks. In addition, the image processing tasks based on deep neural networks have always been a difficult problem. Sometimes a complex model is applied to a large data volume task, often taking days or even weeks to train. In response to these problems, the researchers have also developed a variety of optimization algorithms, designed to improve the convergence speed of the model, and reduce the error of the model, common optimization algorithms such as SGD, Adam and so on. In this paper, we presented a novel architecture that integrating deep energy model into stochastic depth network (SADIE) to build image features. The novel contributions in this paper are as follows: (1) We improved the stochastic depth network by using deep energy model, which can obtain better generative models when training based on the energy model. (2) We applied the Dirichlet distribution related to the dimension of input data to our network when chose whether to skip the layers, since the distribution can improve the performance of the network and make results more accurate. Since the connection has a certain probability of being randomly ignored in the block of the stochastic depth network, the amount of calculation can be reduced. The authors of stochastic depth networks claim that the model after the above adjustments expects a depth of 3/4 and a 40% increase. This feature was also confirmed in our experiments. This design of introducing random variables effectively overcomes the over-fitting to make the model better generalized. Inactivation of a part of the block actually realizes a hidden model fusion. Since the depth of the model is random during training, the depth of the model is determined at the time of prediction, and the information is filtered as the layer is extracted. When the information reaches the upper level of the network, it is not very informative, and the high-level network is difficult to get effective training. We set
doi:10.31614/cmes.2018.04249 fatcat:vytorzfmabhktgocq2mhy7wto4