Real image super resolution network for simplifying the degradation model
Lin Xufeng，Wu Lijun
College of Physics and Information Engineering, Fuzhou University
Abstract： In the task of image super resolution, bicubic down sampling is commonly used to construct datasets for training networks. However, due to the fixed degradation model, bicubic down sampling results in low generalization ability of the network and cannot be used for real world low resolution images. To address this problem, this paper proposes a preprocessing module that combines with the network obtained from the bicubic down sampling dataset to improve its generalization ability while reducing resource consumption. In addition, this paper also designs feature learning training strategies and multi task joint training strategies for different accuracy requirements. By adopting corresponding training strategies according to different requirements, it can meet the accuracy requirements while having the characteristics of low computational resource consumption, fast training speed, and wide applicability. Experiments have shown that adding a network with a preprocessing module can achieve greater improvements in reconstruction effect and perceptual quality with less model parameter increase, and further improve accuracy through different strategies.
Key words : super resolution; preprocessing module; multi task learning; computer vision
引言
單圖像超分辨率（Single Image Super Resolution,SISR）旨在從低分辨率（Low Resolution，LR）圖像恢復高分辨率 （High Resolution，HR）圖像。在訓練SISR的網(wǎng)絡(luò)時，人們常使用二三次下采樣生成超分辨率數(shù)據(jù)集從而使網(wǎng)絡(luò)學習到相應的退化模型，進而恢復圖像高頻分量。但實際低質(zhì)量圖像的形成有兩大主因：成像設(shè)備性能以及環(huán)境因素干擾，這與二三次下采樣生成的低質(zhì)量圖像在退化模型上會有較大出入。學者通過構(gòu)造數(shù)據(jù)集，將真實的LR HR數(shù)據(jù)集應用于超分辨率網(wǎng)絡(luò)的訓練，使超分網(wǎng)絡(luò)能更好地應用于真實的低分辨率圖像。例如利用不同的拍攝器材或調(diào)整參數(shù)構(gòu)造LR HR數(shù)據(jù)集［1-5］以及利用生成對抗模型生成更接近于真實場景的LR HR數(shù)據(jù)集［6］。如圖1所示，與利用二三次下采樣得到的數(shù)據(jù)集不同，真實世界低分辨率數(shù)據(jù)集的退化模型復雜度較高，并且不同的設(shè)備型號以及不同的參數(shù)設(shè)置均會導致退化模型發(fā)生變化。而利用二三次下采樣得到的數(shù)據(jù)集則具有較為固定的退化模型，僅在圖像的高頻分量產(chǎn)生退化，而低頻分量則與原圖近似。
