Image Super-resolution Algorithm Based On Deep Learning

Image super-resolution reconstruction technique is one of the important research topic of machine vision. it has become a hot area of research because of its' wide application value. however, it has brought the huge challenge for super-resolution reconstruction technique because of the quality of reconstructed images, speed and the complex diversity of images. In recent years, the rise of deep learning promotes the development of artificial intelligence and provides a new thought for super-resolution reconstruction technology.Proceeding from the basic theory and summarize of super-resolution, this paper introduces several traditional reconstruction algorithms, which based on interpolation,reconsitution, and learning, according to the classification of super-resolution reconstruction algorithm on the technology and presents the evaluation criterion of reconstructed images. Because the traditional algorithms have some shortcomings in the reconstruction effect and speed, deep learning theory is applied to the reconstruction process of this paper and we do a variety of improvements on this basis. We mainly studied the convolutional neural network's application of image reconstruction, fast visibility restorationy algorithm and image enhancement algorithm.We studied the reconstruction model based on convolution neural network(SRCNN) and another reconstruction model based on fast convolution neural network (FSRCNN) in depth. In view of the deficiency of those models, we propose improvement plan. First of all, Analysing the training process and some parameters of the network. Finally, we put forward a new training strategy and adopt a new activation function(PreLU). The new model greatly reduces the training time and system stability has also been enhanced. Then, according to the problem that the original FSRCNN model is not ideal for the reconstructing of the atomized image,we add the fast visibility restorationy algorithm as the preprocessing process of the model. Finally, in view of the existing models tend to overlook the human visual demand, We add the image enhancement algorithm to the model's subsequent stages.We adopt the modified Retinex algorithm to enhance the overall visual effect of the image, resulting in high-resolution images that are enhanced in detail and more in line with human visual needs.The article combines the subjective experience and the image quality evaluation standard to design comparison experiments and compare reconstruction results.Experiments show that our improved scheme not only has a good reconstruction effect and reconstruction speed for the general image, but also has a great effect on the reconstruction of the atomized image,and the reconstruction process is more stable and efficient. In addition, the image details have been further enhanced and more in line with the human visual demands.