Flow-Guided Deformable Alignment with Channel-Wise Self-Attention Reconstruct for Efficient Burst HDR Restoration

High dynamic range (HDR) imaging algorithms require precise alignment and efficient feature fusion to handle motion and exposure variations. In prevailing HDR reconstruction frameworks, a substantial portion of network capacity is dedicated to feature fusion, whereas alignment modules are typically designed with considerably fewer parameters. This architectural imbalance may hinder the effectiveness of alignment, particularly in handling complex motions, thereby limiting the overall reconstruction quality. To address this issue, we propose an alignment-centric approach that integrates a lightweight fusion module, significantly enhancing alignment accuracy while maintaining computational efficiency. Furthermore, to reduce computational overhead while maintaining alignment robustness, we adopt a representation that decreases spatial resolution while increasing channel dimensionality, effectively preserving essential image information. By doing so, our method not only expands the receptive field without losing critical details but also significantly reduces computational overhead. Extensive experimental results demonstrate that our approach achieves substantial performance improvements with minimal computational cost, securing second place in the NTIRE 2025 Efficient Burst HDR and Restoration Challenge while significantly reducing both parameter count and FLOPs compared to the first-place model. These findings highlight the crucial role of alignment in HDR reconstruction and offer an effective solution for balancing performance and computational efficiency.