Image Super-Resolution Using Convolutional Networks And Its Visual Improvement

Since the super-resolution technology provides high-quality images, and therefore it has been widely used in video, medical and public security systems.Through some algorithms, it can generate the high-resolution images from some low-resolution images. It not only improves the resolution of the image, but also to make up for the hardware problems,such as high cost and to be difficult to achieve. In this thesis, we study to improve the image clarity. The main work of this thesis is the following:(1) Based on the convolutional neural network single Super-Resolution image reconstruction:We will use the full feed-forward neural network convolution, which has three layers, to achieve super-resolution reconstruction for single image. At first, it extracts features; and then through the low-resolution mapping layer into a high-resolution space using the non-linear mapping; and finally by polymerization reconstruct high-resolution images. It directly achieves end to end mapping from the low-resolution image to a high-resolution image with the learning method, and reduces before and after the optimization algorithm. The thesis experiment on its performance and the speed to do trade-offs compared.(2) Unsharp masking techniques:We use unsharp mask algorithms to enhance the reconstruction image. The reconstructed image subtracts the blurred image obtained through the low-pass filtering, to obtain the frequency component. The frequency component image is scaled and added to the reconstructed image. Then it generates an edge enhanced image.(3) No reference image quality evaluation criteria:The image quality evaluation criteria on super-resolution reconstruction method are all full-reference image quality evaluation criteria. But the experiment shows that the full-reference image quality evaluation criteria can not properly represent the true quality of the image. So in this thesis, we reintroduce three no-reference image quality evaluation index, BIQI, NIQE, and SSEQ. At first, BIQI and SSEQ use SVM to recognized the probability of image distortion of each type, based on the characteristics; then SVR calculated for each type of distortion in the image into a corresponding image quality; and finally the type of distortion to the corresponding probability weighted average image quality in order to achieve the final image quality. NIQE is based on differences to measure the quality of the image, which is between fitting parameters from the distorted image and natural images.In summary, this thesis introduces the super-resolution techniques, and introduces non-reference image quality evaluation criteria to assess the reconstruction images and the result of unsharp mask algorithms, while the unsharp mask enhancement technology as the human visual performance optimization. Experiments show that this method not only has lightweight construction, but also can be faster, and get similar or better results than state-of-the-art methods. Unsharp masking techniques enhances the image contour, so that the reconstructed image details more clearly, in order to achieve more consistent with human visual requirements. But when we do experiments found that when a larger noise appears to the image, which needs to reconstruct, yet we can not remove noise clearer, and the visual effects still needs to be improved.