Research On Image Recognition Algorithms Based On Deep Neural Networks

Machine Learning algorithms have important theoretical and practical significance for Artificial Intelligence systems. Over the past decade, represented by Deep Neural Networks, Deep Learning algorithms and models have made remarkable achievements in feature selection and learning tasks. In the fields of image recognition, speech recognition, machine translation and many others, Deep Neural Networks have made many breakthroughs. In some tasks,their data processing capabilities even surpass the artificial level.Deep Neural Network models have shown excellent performances in the tasks of Artificial Intelligence, but there are still some difficult problems remained, such as feature sparsity, scaleinvariant property, crosschannel covariate shift, and frequency domain feature selection ability. Around the above technical problems and based on previous work, we conduct researches and achieved some meaningful results. The main work and innovation of this paper are listed as follows:1. Improve the mask generation strategy of DropConnect algorithm, let the generation function of the mask depends on the output values of the neurons in upper layer, which makes the Neural Network model can dynamically determine the dropping probabilities of neuron connections according to the sparsity of the neuron outputs. The improved model has the ability to select the sparse features. The experimental results show that the sparsity property of the new algorithm is significantly improved by 122.7%. In addition, the recognition accuracy is improved too.2. The scaleinvariant property of Convolutional Neural Network model is studied, and a new scaleinvariant Convolution Neural Network model is proposed, which can automatically adapt to the inplane scale changes of input images. Meanwhile, multilayer Maxout networks are nested into the model in order to improve the ability of feature fitting and extraction. The experimental results show that, compared with the traditional Convolutional Neural Network model, the scaleinvariant property of the new model increases by 8.2%20.1%,and recognition error rate decreases by more than 13.1%.3. In order to solve the problem of crosschannel covariate shift in multichannel Neural Network models, we extend the singlechannel Batch Normalization algorithm to multichannel models, and propose a crosschannel Batch Normalization algorithm. The forward and backward propagations of the crosschannel Batch Normalization algorithm are formulated. Experiments are carried out in Deep Residual Model. Results show that, when using new algorithm,the distributions of neuron activations evolve more stably, and the final recognition error rate decreased by about 4.0%.4. Convolutional Neural Network model lacks of frequency domain feature extraction capacity. We integrate frequency domain filtering operation with convolutional layers, and propose a new Convolutional Neural Network model with the ability to frequency domain feature selection. The new model enhances the ability to extract signal features in particular frequency bands, thereby improves the recognition accuracy. The experimental results show that the recognition error rate decreases from 6.69% to 3.17% in the GTSRB dataset.In summary, this paper researches many problems of Deep Neural Networks such as feature sparsity,scaleinvariant property, and so on. When applied to image recognition tasks, these research results achieved good improvement.