Research On Image Super-Resolution Reconstruction Based On Feature Supplement Module

Image super-resolution aims to reconstruct and supplement the lost information of low-resolution image through effective algorithms,and to obtain high-resolution image with richer information.With the wide application of depth learning technology in super-resolution tasks,the quality of high-resolution images obtained by artificial intelligence algorithms has been far superior to before.Due to the shortcomings of traditional supervised learning in recovering natural images,the methods based on unsupervised learning have emerged.These methods only utilize the given test image.Some of these methods use this image to fine-tune the trained network.These methods have shortcomings in the application of the internal information of the given test image.It is difficult to fully integrate the recovery effect obtained from the benchmark dataset with the recovery effect provided to the network by the given test image.There are also shortcomings in using self-similar features within images.Moreover,the loss function does not pay extra attention to the boundary information that is difficult to recover.To address these issues,the paper adopts the image super-resolution method based on the feature supplement module.The main works are as follows:(1)In view of the low efficiency in fusing different features and high cost of the second iteration of the existing methods,this paper designs a feature supplement module.The feature supplement module is integrated as an independent module in the original network,specifically responsible for extracting the internal features of a given image during secondary training.While the original network obtains initialized parameters through benchmark training dataset,and the parameters do not change during secondary iteration.The experiments have shown that the feature supplement module can improve the fusion efficiency of the model for different features.Because the feature supplement module is a simple structure,the calculation cost of the second iteration can be reduced by fixing the parameters of original network.(2)Aiming at the problem that the feature supplement module fails to make full use of the internal self-similarity characteristics of the given test image,this paper proposes the feature enhancement module.The feature enhancement module calculates the cosine value of the angle between different channels in the form of a vector,as to measure the similarity between the two channels.The similarity coefficient matrix is obtained through the similarity degree of any two channels,and then enhance similar features of each channel in the form of vector product.The experiments have shown that adding feature enhancement module can improve the utilization of similar features within the test image.The performance of the model has been further improved.(3)Aiming at the problem of failing to pay attention to the boundary information at the loss function,this paper proposes the Laplace loss function.The super-resolution image and the real high-resolution image are mapped to other spaces through the Laplace operator,and then calculated the loss between these two images.Adding this loss to the original loss function can increase the weight of boundary information in the overall loss function,as to improve the sensitivity of the model to boundary information.The experiments show that high-resolution images obtained by Laplace loss function have clearer boundary information.The method proposed in this article improves the shortcomings of existing methods in some degree.It improved the use of internal features of a given image in model.Both objective indicators and subjective effects have been further improved.