Feedback Multi-scale Residual Dense Network For Image Super-resolution

Image super-resolution is a technique that reconstructs a low-resolution image into a high-resolution image using some method.Image super-resolution has important applications in many real-world scenarios.In addition,image super-resolution can be used as a powerful pre-processing method to further enhance the original task in other vision tasks.With the application and continuous development of convolutional neural networks in various vision tasks,single-image super-resolution reconstruction based on deep learning has greatly improved the reconstruction of images.However,many current single-image super-resolution reconstruction algorithms based on convolutional neural networks lack multi-scale information and cannot effectively extract high-frequency information from images,resulting in poor quality of reconstructed images,while the network structure generally adopts a feed-forward structure without using feedback information,which helps to gradually refine the image information and obtain more high-frequency information.Based on the above reasons,this paper investigates multi-scale feature extraction and fusion in single-image super-resolution reconstruction,and the utilization of feedback network mechanism,with the following two main parts.(1)A multi-scale feature extraction module that utilize two different types of multi-scale features at the same time is designed to address the problem that existing super-resolution algorithms utilize a single type of multi-scale feature when using multi-scale features.A feature fusion method based on an error feedback mechanism is proposed for the problem of inefficient and ineffective multi-scale feature fusion in existing single-image super-resolution algorithms.A feedback multi-scale dense residual network suitable for image super-resolution reconstruction is designed by combining the above two schemes.The network utilizes two different multi-scale features simultaneously,one for multi-scale features with different perceptual fields extracted using convolutional kernels of different sizes,and the other for multi-scale features of different sizes.The proposed network is compared with other methods at different scaling ratios and better results are obtained in both subjective and objective evaluation metrics.(2)The proposed feedback multi-scale dense residual network and other current single-image super-resolution algorithms suffer from lack of effective feedback connections,underutilization of features,and excessive model size,and a global feedback multi-scale dense residual network for image super-resolution algorithm is designed to address this problem.In addition to using an error feedback mechanism in feature fusion,a global feedback connection is used to improve feature utilization,progressively refine features,and also effectively reduce the number of parameters in the model and have better reconstruction quality.The global feedback connection re-inputs the multi-scale features extracted by the cascaded extraction modules into the front-end to achieve global feature feedback,and also achieves effective reduction of the model size by sharing the module parameters and using multiple global feedback loops to achieve progressive refinement of the features and obtain more high-frequency information.It is experimentally demonstrated that the proposed global feedback network achieves better reconstruction results at different scaling scales than other single-image super-resolution methods and the proposed feedforward network.