Main Image Enhancement Based High Dynamic Range Imaging Using Dual-Lens Systems

With the booming development of mobile devices,dual-lens systems are widely deployed in high-end smart phones.After letting the dual-lens shoot images in the same shot-time,the shot images can be fused to obtain the output image with higher quality.In daily life there are some situations where the dynamic range of the scene is very high,such as backlighting scenes.In this case,to get a higher quality image,we need to use high dynamic range(HDR)imaging technology.And HDR imaging is one of the promising applications that the dual-lens system can outperform largely than the traditional single-lens system.This thesis uses convolutional neural network(CNN)to propose a dual-lens HDR imaging model based on the main image enhancement strategy.The short-exposure image is used as the main image and the long-exposure image is used as the reference image.The main image is enhanced through soft exposure,image alignment,guided denoising,and image fusion to generate a ghost-free HDR image.Since there is a lack of datasets for such problems in the field,this thesis builds a new dataset using a dual-lens system.The main work of this thesis is as follows:(1)This thesis uses a dual-lens system and shoots the same scene with different exposure to build an HDR imaging dataset.The work includes dual-lens system construction,dual-lens system calibration,data capture,and ground truth generation.(2)This thesis proposes a new algorithm for HDR imaging based on main image enhancement.Instead of merging the multiple low dynamic range inputs directly,this thesis uses an indirect way that enhances the main image,i.e.the short exposure image,using the long exposure image serving as guidance.In detail,this thesis proposes a new model which consists of three subnets,i.e.Soft Warp CNN,3D Guided Denoising CNN and Fusion CNN.(3)This thesis proposes a 3D Guided Denoising CNN to learn the filtering weights with context for generating spatially consistent denoising results.Each pixel of the denoising result generated by this 3D Guided Denoising CNN is obtained by the weighted average of the neighboring pixels.(4)The proposed algorithm is compared with the state-of-the-art HDR imaging methods.Experimental results show that the proposed algorithm outperforms related methods largely and avoids ghost artifacts.