HDR Image Reconstruction Based on Asymmetric Non-local Residual Attention Network Utilizing Neural Networks to Eliminate Ghosting in HDR Image Reconstruction of Dynamic Scenes

High Dynamic Range (HDR) image reconstruction technology can fuse multiple Low Dynamic Range (LDR) images captured at different exposure values into a single HDR image. However, in dynamic scenes, mitigating or removing ghosting and artifacts caused by the movement or misalignment of objects in the reconstructed image presents a critical challenge. Recent reconstruction methods based on Convolutional Neural Networks (CNN) perform local attention-based fusion of multiple LDR images to reconstruct HDR content, but still struggle to effectively eliminate ghosting and artifacts. In our proposed ANRANet, residual attention network efficiently and accurately reconstructs HDR images. To alleviate ghosting and artifacts, the network employs a feature processing block with nested residual learning (Residual-in-residual learning) to pass low-frequency information and refines and optimizes the generation of latent representations through a global context attention mechanism. Then, by fusing different levels of features while considering long-range dependencies through an asymmetric non-local block, it better captures the correlation information between each pixel. Experimental results show that the proposed model achieves excellent performance; details in areas of abnormal exposure are effectively restored, and ghosting and artifacts caused by significant motion of objects across different exposures are eliminated. Compared to existing methods, our method achieves higher results in both objective and subjective evaluations.