Research And Implementation Of Single Image Dynamic Range Enhancement Based On Lightweight Convolutional Neural Networks

Due to the limited dynamic range of a standard digital camera,it quintessentially fails to store all dynamic range of a scene.Consequently,the captured images,namely low dynamic range(LDR)images which often contain under-exposure and over-exposure regions simultaneously,Nowadays,technological changes make people become more and more picky about image quality,traditional LDR images have gradually failed to meet the requirements,and high dynamic range(HDR)imaging technology has emerged as the times require,the technology traditionally makes use of multiple LDR photographs with different exposures and then combine these photographs into a single HDR image.However,most existing images are not captured at different exposures,how to reconstruct these LDR images into corresponding HDR images to improve the visual quality has become a popular research topic.The algorithms of reconstructing an HDR image from a single exposed LDR image is commonly referred to as inverse tone mapping operators(iTMOs).Retrieving the lost information in the LDR image is a typical ill-posed problem,most recently proposed methods thus make use of complex convolutional neural networks(CNNs)to generate an HDR image from a single LDR image.However,these methods require high computing power and thus are too cumbersome to run on mobile devices which quintessentially have limited computational resources,restricting their potential application.Perceptual experiments have shown that the perceived quality of an image degrades substantially with over-exposure.To this end,this thesis proposes a lightweight CNN,namely LiTMNet which takes a single LDR image as input and recovers the lost information in the saturated regions of the image to reconstruct an HDR image.LiTMNet is essentially a symmetric encoder-decoder structure equipped with skip-connections.It contains a both effective and efficient upsampling block which does not only improve the quality of the reconstructed HDR image,but also significantly accelerates the reconstruction.The final HDR image is produced by nonlinearly blending the network prediction and the original LDR image.Qualitative and quantitative comparisons demonstrate that LiTMNet produces results of high quality comparable with the current state of the art and is 37× faster as tested on a mobile device.LiTMNet improves the flexibility of deep iTMOs significantly since it can be introduced to mobile devices to provide end-to-end HDR reconstruction.Owing to the fact that a smartphone can equipped with 12-megapixel or even higher sensor,how to enable high resolution input is a common problem shared by CNN-based algorithms,and such a problem is more challenging on mobile devices.To tackle this problem and further improve the practicality of LiTMNet,we combine it with guided upsampling and build an Android app.I.e.,given a high resolution input,the app first downsamples the input,then process it using LiTMNet,finally it applies guided upsampling to the prediction of LiTMNet to get the high resolution output.However,using an overexposed LDR image as guidance map can lead to unwanted artifacts,to address this problem,we present an efficient method of generating the guidance map.In this manner,the app is able to reconstruct a 12-megapixel LDR input within 3 seconds and display the result after tone mapping.