Ghost-Aware Dual-Branch High Dynamic Range Imaging For Dynamic Scene

In the study of high dynamic range (HDR) image reconstruction from multiple low dynamic range (LDR) images, the challenge of resolving ghosting artifacts caused by moving objects is a significant hurdle. To overcome this challenge, we propose a ghost-aware dual-branch deghosting algorithm tailored for dynamic scenes in HDR imaging. Our algorithm employs a parallel dual-branch feature extraction network, efficiently extracting features at both local and global scales. For local features, we have designed a Ghost-Aware Attention Module (GAAM) to capture details and detect and eliminate ghosting artifacts within local ranges caused by motion. For global features, we utilize an Asymmetric Pyramid Non-Local Block (APNB), which captures long-range dependencies between different regions of the image by correlating the entire feature map, compensating for information loss due to extensive motion and removing ghosting artifacts. Furthermore, we leverage a gradient loss function that uses image gradient information to further constrain the network, enhancing details and removing ghosting artifacts, thereby achieving high-quality HDR image reconstruction. Through experiments on a publicly available dataset specifically designed for HDR imaging, we have validated the exceptional performance of our proposed method in handling ghosting artifacts.