High dynamic range image reconstruction based on efficient multi-scale attention

In real-world scenarios, ghosting artifacts caused by object motion and the high computational complexity of traditional transformers have persistently challenged high dynamic range (HDR) image reconstruction. To address these issues, we propose an HDR image reconstruction network integrating a hybrid perception transformer with efficient multi-scale attention mechanisms. Specifically, we design a spatial attention module to achieve cross-exposure feature fusion and construct an efficient hybrid fusion module that synergistically enhances global-local feature interaction through an efficient context extractor and the window attention mechanism. Experimental results demonstrate that our proposed network effectively preserves texture features from differently exposed images while reducing information interference in over-/under-saturated regions, achieving favorable reconstruction performance with enhanced computational efficiency.