Ghost-free high dynamic range imaging with shift convolution and streamlined channel transformer

Abstract

High dynamic range (HDR) imaging merges multiple low dynamic range (LDR) images to generate an image with a wider dynamic range and more authentic details. However, existing HDR algorithms often produce residual ghosts due to challenges in capturing long-range dependencies in scenes with large motion and severe saturation. To address these issues, we propose an HDR deghosting method with shift convolution and a streamlined channel Transformer (SCHDRNet). Specifically, to better aggregate information across frames, we propose a pixel-shift alignment module (PSAM) to enhance the interaction of adjacent pixel features through shift convolution, improving the accuracy of the attention alignment module (AAM). Additionally, we propose a hierarchical streamlined channel Transformer (SCT) that integrates streamlined channel attention, multi-head self-attention, and channel attention blocks. This architecture effectively captures both global and local context, reducing ghosting from large motions and blurring from small movements. Extensive experiments demonstrate that our method minimizes ghosting artifacts and excels in quantitative and qualitative aspects.