Hierarchical Adaptive Attention and Multi-Scale Transformer for Ghost- Free HDR Imaging

Abstract High Dynamic Range (HDR) imaging synthesizes vivid images by merging multiple low dynamic range (LDR) images of different exposures. Nevertheless, in dynamic scenes, object motion or camera commonly introduce ghosting artifacts, severely degrading image quality. Although numerous DNN-based methods have been proposed to address this issue, existing solutions remain unsatisfactory. Spatial attention-based approaches often struggle to cope with complex scenarios characterized by random luminance fluctuations and large-scale motion, while conventional HDR deghosting models that rely on CNN during the fusion stage are hampered by limited receptive fields, lack of dynamic weighting and the absence of multi-scale capabilities.To overcome these limitations, we propose two innovative modules. The Luminance Adaptive Channel Attention (LACA) module dynamically and adaptively modulates channel- wise weights across multiple scales. This enables precise information balancing among channels, effectively suppressing ghosting artifacts and alleviating color saturation issues, thereby yielding refined feature representations that enhance the HDR fusion process.The Multi-Scale Residual Swin Transformer Block (MSRSTB), empowered by a multi-scale Transformer architecture, provides an expansive receptive field and dynamic weighting mechanism. It adeptly handles diverse motion patterns, integrating features in a hierarchical, coarse-to-fine manner, and efficiently manages regions with varying exposure levels. As a result, it significantly reduces saturation artifacts and mitigates ghosting, facilitating high-quality HDR image reconstruction in challenging scenarios. Comprehensive qualitative and quantitative evaluations emonstrate that our proposed modules outperform state-of-the-art methods.