Research On The Improvement Of Deep Learning Network Based On Visual Physiological Mechanism

Deep Learning technology is a branch of traditional machine learning, and it has made remarkable achievements in decade. By combining of multiple nonlinear feature extractors, Deep Learning technology can transform the original data to a higher dimension, so that we can find the distributed features in these data. Deep Learning technology can generate feature extractors automatically, which overcomes the disadvantages of traditional machine learning algorithms. Nowadays, Deep Learning technology has been widely used in computer vision, natural language process, etc.Deep Learning technology improved people's living standards greatly.In Deep Learning technology, the model commonly used to process image data is the Convolutional Neural Networks (CNNs). CNNs extract the high-level feature at different level of the image by using multi-layer convolution operation, and these features can be used in further tasks such as classification and segmentation. This process is similar to the process of visual processing in human visual cortex. Typically,it takes researchers' a lot of time to design a Deep Learning model and to write code,which has a significant negative effect on research efficacy. In this thesis, we have chosen the mature and widely used deep learning framework named Caffe, which is specific for convolutional neural networks. With the advantages of expression, speed,and modularity provided by this framework, researchers could quickly achieve the experimental paradigm.Based on the Caffe framework, following work and research have been done in this thesis:Firstly, introduce the research status of deep learning technology in image recognition and natural language processing in recent years, and briefly introduces the hardware and software dependence of deep learning technology and introduce the Caffe framework as well.Secondly, combined with the neural mechanism model of human visual information processing and the neural network model in machine learning, we introduce the traditional shallow learning model named perceptions. In addition, some deep learning technologies are presented as well. At the end of this chapter, the principle and structure of the convolution neural network that is widely used in the field of image processing is introduced emphatically.Again, by visualizing the weights in post-trained CNNs, we can figure out that these geometric figures have obvious physiological significance. In this thesis, we improve the performance of the CNNs by using the Gabor filter to modify the low-level convolution kernel in CNNs, and quantitatively analyze the difference between the modified model and the original model on MNIST dataset and the CIFAR10 dataset.Then, since the gray point existing in the color-biased image have rich statistical information, which can be used in correcting the color-biased image. So we designed a Convolutional Neural Network to predict whether or not each pixel in the color-biased image is a gray point and to help the color constant problem. The whole framework have achieved a certain effect.Finally, a brief review of the work of the master period was presented. I summed up the lack of research and expectations.