Attention Based Convolutional Neural Networks For Super Resolution

Image plays an indispensable role in human life,but there exist many restrictions which heavily influence image quality,and the resolution of image may not satisfied our requirements.This not only causes a bad experience for people,but also affects objective tasks,such as image understanding and image classification.Super-Resolution aims at breaking this limitation by using one or more low-resolution images to reconstruct corresponding high-resolution images through software algorithms.This paper mainly discusses and analyzes deep learning-based super-resolution methods,and proposes three super-resolution algorithms which combined the convolutional neural network with the attention mechanism.The works are summarized as follows:(1)We present a novel densely convolutional attention network for single image superresolution(SISR).The existing residual structure treat features equally across channels,which may reduce the representational ability of networks.In order to solve this problem,we propose a convolutional attention mechanism which adds an weighted skip-connection joint the original convolution.Our network can adaptively recalibrate convolutional-wise features by considering interdependencies among convolutional layers.Besides,the proposed network leverage several dense block to enhance the relationship between different convolutional layers,more information could be used sufficiently by this block learning method,in order that reconstructed images can obtain richer details and clearer edges.Experiments show that our densely convolutional attention network is superior to several state-of-the-art methods in both quality and quantity.(2)We propose a non-local hierarchical residual network for SISR.Most deep learning based methods lack the ability to distinguish features in network.For image super-resolution,it is important to design an effective prior to learn the correlation of various feature.To resolve the problem,we employ the non-local attention module to measure the self-similarity among feature maps by convolution and matrix operations.It captures the global information from features and enables the network to obtain a priori information related to low-resolution images.Thus our method reconstruct images with a sharper edge.In addition,we employ group convolutions to build a hierarchical residual structure,which enable the network extract image features hierarchically.It can reduce the executive time while ensuring reconstruct image quality.Experimental results show that our method achieves better performance and speed against state-of-the-art methods.(3)Low-Quality video usually has complex distortions,such as low resolution,noise and abnormal exposure.We design a super-resolution algorithm for low-quality videos by analyzing the characteristics of them.We explore video compression processing to extract video key frames,building a training set that minimize redundant data while maintaining the diversity of video content.The algorithm combines multi-scale attention mechanism and the residual in residual module to construct a two-stage network.Video frames are pre-processed through the color-correction stage for color correction and noise removal,and then enlarged in the super-resolution stage.All low-quality video frames are processed by the network during test phase,and aggregated into high-quality video.Experiments show that the algorithm can recover low-quality videos of different types and different contents.Our algorithm utilizes less resources and has excellent performance.