A Novel High Dynamic Image Fusion Method via an Unsupervised End-to-End Framework

For the sake of high-quality images of the high dynamic range (HDR) scenes, it is effective means to fuse the multiexposure sequences for the same HDR scene. However, the fused images using the existing fusion methods are prone to detail loss or block effect. Aiming at these problems, a novel unsupervised end-to-end framework is developed to provide solutions for the multiexposure image fusion. Instead of conventional manual setting, the optimal image weight coefficients of the multiexposure images are learned automatically, which makes this model more suitable for application. Most importantly, a customized loss function is designed to enhance the network achievement and automatically learn the parameters in the direction of optimal fusion image. According to the quantitative and qualitative results of a large number of experiments, it is demonstrated that the proposed framework performs its superiority and effectiveness compared with the state-of-the-art approaches.