Bidirectional Image-Event Guided Fusion Framework for Low-Light Image Enhancement

TL;DR

BiLIE tackles extreme low-light image enhancement by fusing frame-based images with event data through two key components: Dynamic Adaptive Filtering Enhancement (DAFE) and Bidirectional Guided Awareness Fusion (BGAF). A high-quality RELIE dataset is introduced to provide faithful ground-truth references and robust evaluation. Empirical results show BiLIE achieving state-of-the-art performance on both RELIE and LIE, notably a PSNR gain of 0.81 dB on RELIE, with improved edge preservation and color fidelity. This work enhances multimodal low-light enhancement by enabling mutual, structure-aware guidance across image and event modalities, with practical implications for surveillance and autonomous systems.

Abstract

Under extreme low-light conditions, frame-based cameras suffer from severe detail loss due to limited dynamic range. Recent studies have introduced event cameras for event-guided low-light image enhancement. However, existing approaches often overlook the flickering artifacts and structural discontinuities caused by dynamic illumination changes and event sparsity. To address these challenges, we propose BiLIE, a Bidirectional image-event guided fusion framework for Low-Light Image Enhancement, which achieves mutual guidance and complementary enhancement between the two modalities. First, to highlight edge details, we develop a Dynamic Adaptive Filtering Enhancement (DAFE) module that performs adaptive high-pass filtering on event representations to suppress flickering artifacts and preserve high-frequency information under varying illumination. Subsequently, we design a Bidirectional Guided Awareness Fusion (BGAF) mechanism, which achieves breakpoint-aware restoration from images to events and structure-aware enhancement from events to images through a two-stage attention mechanism, establishing cross-modal consistency, thereby producing a clear, smooth, and structurally intact fused representation. Moreover, recognizing that existing datasets exhibit insufficient ground-truth fidelity and color accuracy, we construct a high-quality low-light image-event dataset (RELIE) via a reliable ground truth refinement scheme. Extensive experiments demonstrate that our method outperforms existing approaches on both the RELIE and LIE datasets. Notably, on RELIE, BiLIE exceeds the state-of-the-art by 0.81dB in PSNR and shows significant advantages in edge restoration, color fidelity, and noise suppression.

Figures (11)

• Figure 1: The Enhancement results of the frame-based method cai2023retinexformer, image-event fusion-based method liang2024towards and the proposed method on LIE and our constructed RELIE. Comparing (c) and (d), our constructed dataset exhibits higher image quality. The second row demonstrates that our method effectively suppresses noise and artifacts, resulting in higher-quality enhanced images. Benefiting from our Dynamic Adaptive Filtering and Bidirectional Guided Awareness fusion design, along with the use of frequency loss, our method effectively suppresses the noise and artifacts introduced during ground truth construction, resulting in smoother and more natural visual outputs.
• Figure 3: The network architecture of our proposed method. BiLIE adopts a dual-branch encoder-decoder structure, consisting of two fundamental units: Dynamic Adaptive Filtering Enhancement (DAFE) and Bidirectional Guided Awareness Fusion (BCAF). These units work together to generate high-quality, smooth images with clear contours.
• Figure 4: The structures of the Learnable Filter Block (LFB) and the Learnable Gating Block (LGB), as well as the feature visualization results of the Fixed Filter Branch within the Dynamic Adaptive Filtering Enhancement (DAFE) module. $F_{event}$, $F_{shift}^{'}$, $F_{event}^{1}$, and $F_{E}^{1}$ respectively represent the feature map before DAFE processing, after Fourier transform and frequency-domain shifting, after the high-pass filter, and after DAFE processing. Compared to the original event representation $F_{event}$, the filtered result $F_{E}^{1}$ is more natural and sharper.
• Figure 5: Motivation analysis and feature visualization for bidirectional image–event guidance.
• Figure 6: Qualitative results on indoor scenes from the RELIE dataset.