Research On Image Super-Resolution Algorithm Based On Deep Learning

Digital images are one of the main communication media in the era of intelligent media.However,since images suffer inevitable degradation during acquisition,compression,and transmission,the quality of images acquired in the real world is degraded.Image superresolution technology is a classic low-level task in computer vision and image processing,aiming to reconstruct high-resolution images with high-frequency details from corresponding low-resolution images.With the rapid development of deep learning technology,convolutional neural network-based super-resolution algorithms have been widely used in various research fields,such as medical imaging,remote sensing imaging,public safety,autonomous driving,intelligent display,etc.Although the convolutional neural network-based image super-resolution algorithms have achieved advanced experimental performance,there are still many challenging problems to be solved in practical applications,such as low prediction accuracy of blur kernel and degradation inconsistency in unguided image super-resolution tasks,poor guide information transfer in guided image super-resolution tasks,etc.In view of the above problems,based on the powerful learning ability of the convolutional neural networks,the research and exploration of image super-resolution algorithms are carried out.The specific research contents mainly include:On the one hand,the existing blur kernel estimation algorithms in the unguided image super-resolution tasks need to iteratively optimize the training in the inference phase,which is time-consuming;on the other hand,the model is optimized based on the low-resolution image space domain,which makes the accuracy of predicting the blur kernel is poor.In view of the above problems,a blur kernel estimation network based on blur kernel space domain optimization is proposed to reduce the estimation error caused by indirect optimization.In addition,the network model is trained to optimize the model parameters based on an adaptive attention loss function,which increases the attention to important regions in the blur kernel space.Experimental results show that the proposed blur kernel estimation network can not only obtain accurate blur kernels but also achieve state-of-the-art blind image super-resolution performance by combining with image super-resolution network models for various degradation.The peak signal-to-noise ratios of kernel estimation and super-resolution reconstruction are improved by at least 13.58 dB and 0.89 dB,respectively.Most of the existing convolutional neural network-based super-resolution algorithms assume that the degradation is known and fixed,such as bicubic downsampling.However,the performance of algorithmic models suffers from a severe drop when the actual degradation mismatch the training one.To solve the above problems,a blind image superresolution network based on a prior correction network is proposed,causing the superresolution network based on a single fixed blur kernel to be also applicable to other unknown blur kernels.Specifically,the prior correction network consists of a blur kernel estimation network,a correction filter module,and a correction refinement network.The blur kernel estimation network estimates the unknown blur kernel from the input low-resolution image,the correction filter module corrects the input image according to the estimated blur kernel to make it adapt to a specific degradation distribution,and the correction refinement network adjusts the filtered output image to eliminate the influence of blur kernel mismatch or misestimate.Experimental results on different datasets show that the proposed prior correction network combined with image super-resolution algorithms for single degradation achieves state-of-the-art blind image super-resolution performance,and its peak signal-tonoise ratio is improved by at least 0.72 dB.Most of the current mainstream algorithm models in guided image super-resolution tasks unilaterally transfer the high-frequency structural information in the guided image to the target image,ignoring the correspondence between the guided image and the target image.To address this problem,a guided super-resolution algorithm based on a dual autoencoder attention network is proposed.The algorithm model mainly includes two subnetwork models,where the guided autoencoder network is used to reconstruct highresolution guided RGB images while extracting high-frequency structural features and transferring them to the target auto-encoder network;the target auto-encoder network fuses high-frequency structural features and reconstructs high-resolution depth maps.Specifically,all auto-encoder networks adopt similar structures and are trained simultaneously to ensure structural consistency.Furthermore,an attention mechanism is introduced into the model to enhance the feature extraction of high-frequency edge regions.Experimental results show that the experimental performance of the proposed dual auto-encoder attention network achieves state-of-the-art guided super-resolution reconstruction performance with at least a 0.17 reduction in root mean square error.