Research On Learning-based Image Super-resolution Algorithms

Image super-resolution is a research hotspot in the field of image processing in recent years. Image super-resolution technology use single or multiple low resolution images,digging the correlation and complementary information in low resolution images, improve the spatial resolution of the target image in the form of software algorithm.Through the image super-resolution technology, we can use the low resolution imaging system get higher resolution images in a way that is flexible and low cost.Due to the introduction of training library information, the methods based on learning has a greater potential to improve the resolution of the image compared with the algorithms based on interpolation and based on the reconstruction.In this paper, we focused on the problem of image super-resolution and proposed two kinds of learningbased image super-resolution method: method based on locally linear embedded and method based on convolutional neural network.Image super-resolution method based on locally linear embedded introduces the locally linear embedding algorithm which is a classical method in manifold learning into image super-resolution problem.In this method,we use the locally linear embedded algorithm to dig the relationship between low resolution image block and high resolution image block, and then reconstruct the high resolution image correspond to the target low resolution images.On this basis, this paper proposed a similar block search strategy which uses four angles and two mirrors to improve the effect of super-resolution,and designed a set of experiments to verify the validity of the algorithm.Image super-resolution method based on convolution neural network applied the convolutional neural network model which is a very popular deep learning method in recent years to image super-resolution problem.we first build a convolutional neural network specially designed for image super-resolution.And then we trained the network use a lot of low resolution image block with the corresponding high resolution image as the training samples and the stochastic gradient descent method as training method. After sufficient training and optimizing, the network then represents a a mapping relationship of low resolution image to high resolution image.At this time,we put a low resolution imageinto the network as input image,then we can get the corresponding high resolution image.