Image Restoration Tasks Based On Deep Learning

In the information age of today,with human subjective demand for high quality images and video,image and video processing is becoming more and more widely used in various fields.Image Restoration technology can significantly improve the image or video quality under the existing hardware conditions,and effectively recover the details of the target scene.In recent years,due to the strong self-learning ability,deep learning can end-to-end learn the feature maps of different quality spaces,bringing a new development concept for the field of image restoration.In the track of image restoration tasks,existing most deep convolutional neural networks aim at improving the performance of single image restoration tasks,and cannot properly deal with multiple restoration tasks.An ideal image restoration network should be able to support multiple restoration tasks well.At the same time,the different selection of feature scales is crucial to extracting effective information from different image restoration tasks.To this end,we propose the cross-scale dense residual network(CDRN)to exploit scale-related features and the inter-task correlations among the three tasks.The proposed network can extract multiple spatial scale features and establish multiple temporal feature reusage.The network contains cross-scale dense residuals blocks to adaptively learn different scale features;and two different connection methods: intra-block cross-scale connections and inter-block dense connections,to enhance feature fusion rate and reuse rate.We apply CDRN to three image restoration tasks,i.e.,image denosing,super resolution and JPEG deblocking.Comprehensive experiments demonstrate the necessity of the CDRN and its unanimous superiority on all three tasks over the state of the arts.In the track of video super-resolution reconstruction tasks,current methods based on deep learning usually fail to balance the inter-frame spatial relationship and the intra-frame time relationship,thus ignoring certain crucial inter-frame information from the original low-resolution frame sequence or the hierarchical features of deep networks.In this paper,we propose a novel method for video super-resolution,called dense-connected residual network(DCRnet),to address the above drawbacks.The DCRnet can preserve the low-frequency contents of motion-compensated frames,and facilitate the restoration of the high-frequency details by exploiting the hierarchical features of all the convolutional layers.Specifically,we propose a dense-connected residual block(DCRB)as the basic component,which explicitly mine long-term memory through an adaptive learning process.Extensive experimentation demonstrates that our method is superior to the current state-of-the-art methods in both quantitative and qualitative metrics.