Improving Bracket Image Restoration and Enhancement with Flow-guided Alignment and Enhanced Feature Aggregation

TL;DR

BracketIRE addresses HDR reconstruction from noisy, motion-blurred, multi-exposure RAW inputs by proposing IREANet, which combines flow-guided alignment (FFAM) with enhanced feature aggregation (EFAM) built on the Enhanced Residual Block (ERB). Bayer-preserving augmentation (BayerAug) improves generalization while maintaining RAW Bayer patterns. The approach yields state-of-the-art results on the NTIRE 2024 BracketIRE dataset, with notable PSNR, SSIM, and perceptual gains, and ablation studies confirm the contributions of FFAM, EFAM, and BayerAug. This work advances robust multi-exposure HDR reconstruction by improving alignment, fusion stability, and detail preservation in challenging real-world captures.

Abstract

In this paper, we address the Bracket Image Restoration and Enhancement (BracketIRE) task using a novel framework, which requires restoring a high-quality high dynamic range (HDR) image from a sequence of noisy, blurred, and low dynamic range (LDR) multi-exposure RAW inputs. To overcome this challenge, we present the IREANet, which improves the multiple exposure alignment and aggregation with a Flow-guide Feature Alignment Module (FFAM) and an Enhanced Feature Aggregation Module (EFAM). Specifically, the proposed FFAM incorporates the inter-frame optical flow as guidance to facilitate the deformable alignment and spatial attention modules for better feature alignment. The EFAM further employs the proposed Enhanced Residual Block (ERB) as a foundational component, wherein a unidirectional recurrent network aggregates the aligned temporal features to better reconstruct the results. To improve model generalization and performance, we additionally employ the Bayer preserving augmentation (BayerAug) strategy to augment the multi-exposure RAW inputs. Our experimental evaluations demonstrate that the proposed IREANet shows state-of-the-art performance compared with previous methods.

Figures (6)

• Figure 1: Comparisons between the propsoed IREANet and other representative methods yan2019attentionliu2022ghostzhang2024bracketing. Our result is free of noise and blur while producing more details.
• Figure 2: Illustration of the overall framework of the proposed IREANet. As shown in Fig. \ref{['fig:pipeline']}(a), the pipeline mainly consists of four components: feature extraction, feature alignment, feature aggregation, and reconstruction. Fig. \ref{['fig:pipeline']}(b) illustrates the proposed flow-guided feature alignment module (FFAM). Fig. \ref{['fig:pipeline']}(c) depicts the differences between the vallina residual block (RB) and the proposed enhanced residual block (ERB).
• Figure 3: Illustration of the adopted Bayer preserving augmentation (BayerAug).
• Figure 4: Qualitative Comparison with existing state-of-the-art methods AHDRNet yan2019attention, CA-ViT liu2022ghost, and TMRNet zhang2024bracketing. Our approach can effectively restore high-quality image details from multiple exposure inputs.
• Figure 5: Qualitative results of our ablation study on the Bayer preserve augmentation.