Research And Application Of Image Super-resolution Algorithm Based On Dense Residual Network

With the development of display components,images have gradually become one of the main sources of information for human beings and one of the most commonly used information carriers in human social activities.Image resolution is an important indicator for evaluating image quality,and to a large extent represents the richness of the information contained in the image.Generally,higher image resolution means more information can be captured by the human eye.However,in real life,due to image coding(in image transmission,transmission and storage),limitations of imaging equipment and imaging environment,etc.,images often suffer from varying degrees of texture detail and edge detail loss,resulting in image quality.decline.Super-resolution reconstruction technology can effectively improve such problems,and because this is a software-based method,it can improve the image resolution without using new hardware and reacquisition.However,existing super-resolution algorithms usually require a very deep network structure and a long training time.In addition,current super-resolution algorithms do not treat low-,medium-,and high-frequency information in images separately.This limits the learning ability of these algorithms.In addition,the model of the previous algorithm can only complete an image super-resolution with a fixed scale factor,so multiple models must be trained.In order to try to solve the above problems,this paper proposes an improved method MDLN(Meta-Upscale Densely Residual Laplacian Network)based on dense residual network.The main work of this article is as follows.(1)Feature extraction part: This paper introduces the Laplacian attention mechanism in the dense residual block,which can adaptively rescale the features and build a featuredependent model for learning features at multiple frequencies.Then improve to get a dense residual Laplacian module containing three components of residual block unit,compression unit and Laplacian pyramid.And the design of cascading residuals is used between this new module to improve the efficiency of the deep network and make the network learn more accurate features.(2)Feature amplification part: This article makes full use of the role of meta-learning in neural networks.Successfully borrowed the Meta-Upscale module proposed by Hu et al.In Meta-SR,dynamically predicted the weight of the amplification filter according to the input scale factor,and then generated HR images of arbitrary size according to these weights.Achieve super-resolution tasks with multiple scale factors for a single model.(3)In order to put the improved MDLN algorithm into practical application,this paper designs a super-resolution prototype system with GUI interface.Used to display the comparison between the input image and the super-resolution image,and the evaluation results on each test set.After a series of comparative experiments,it is fully proved that the algorithm proposed in this paper is better than the basic dense residual network to a certain extent,and at the same time improves the above-mentioned problems in the previous superwresolution algorithm.