Deep convolutional networks super-resolution method for reconstructing high frequency information of the single image
Mandan Zhao, Chuanqi Cheng, Zhenjie Zhang, Xiangyang Hao · 2017
Super-resolution image reconstruction is one of the important issues in the field of computer vision. Machine learning is also the powerful method to solve the problem of computer vision. The method of SRCNN, which is put forward by Tang, is used for image super-resolution and shows the state-of-the-art performance. However, the method of SRCNN still has some shortcomings. On one hand, the training network convergence is slow and one trained network can only handle one settled scale factor model. Meanwhile, the input and output of the network contains a large number of the same information. It could lead to redundant training. On the other hand, the accuracy of the super-resolution reconstruction needs to be further improved. In order to improve these weaknesses and inspired by the network of the VGG and ResNet, we propose the method to restore the high frequency information of the image based on the residual network and enlarge the range of the effective vision perception based on very deep convolutional network. Through these improvements, we speed up the convergence rate of the training and improve the super-resolution image reconstruction accuracy. Especially some areas such as image detail and depth discontinuity have good performance. These improvements can furtherly optimize the single image super-resolution method, and expand its application scope. In the end, comparing with several advanced methods, such as SRCNN though the experiment, we confirm the validity and practicability of the improved strategy.