Image Super-resolution Algorithm Based On Deep Learning

In our social life, the image is the most popular information carrier. However, in the process of generating image, the final image can not meet people's requirements due to the image degradation factors, which is an obstacle to people's understanding of the world, and to understand the laws of society. And for how to improving the image quality, the improving of the clarity of the image becomes a subject of the field of image processing. The technology of image super-resolution comes into being because of the above background. And it is successfully applied to the computer vision, medicine, transportation, face recognition and other fields, which has drawn attention of all walks of life. For the method of super resolution, the cost of the high precision image extraction based on the hardware is too high, so the technology of super resolution based on the software became the effective means. In particular, super resolution based on learning becomes a hot subject in the field of image restoration.With the arrival of the era of big data, benefit from increase in available computing power and data increase, the deep learning concept was proposed by Professor Hinton in 2006 has also been great development. Deep learning is a way of learning by imitation of the human brain to establish network, and the complex and structure characteristics of high dimensional data can be found. In the field of image it also has amazing performance, and in many fields win over the other machine learning techniques. In view of the achievements of the deep learning in the major fields, we introduce the deep learning into the image super resolution. In this paper, the article put forward two image super-resolution algorithms based on deep learning, which can be divided into the following:(1) An image super resolution algorithm based on multi-layer perceptron is proposed. By using the depth structure of the multi-layer perceptron to identify the features of the high dimensional image, and thus the super resolution model is established. On the efficiency of time, the training time of the deep neural network for super resolution is far less than the general algorithm, with a good time efficiency. For the effect of image restoration, benefit from increase in available computing power and data increase, the longer the time, the larger the data set, the better the effect will be achieve. After the network training, we will put the picture into the network for super resolution, and get a super resolution image. Experimental results show that, Image super-resolution algorithm based on multi layer perceptron has achieved good results.(2) An image super resolution algorithm based on convolutional neural network is proposed. Convolutional networks and multi-layer perceptron are both a feedforward neural network. The concept of receptive field(i.e., weight sharing) greatly reduces the parameters of the network, reduces the complexity of the model, and the layout is more close to the human brain model. We us three convolution layer for image processing algorithm, has not the pool layer, through the network from the sub-image in MATLAB were superimposed and weight matrix were division processing, the final get the image super resolution image. Finally, the experimental results of the super resolution images and the parameters of the network are analyzed in the article. Experimental results show that, Image super-resolution algorithm based on CNN has achieved good results.