RLFED-NET: A Robust Network for Feature Extraction and Descriptor Computation in Low-Light Scenarios

Feature matching fundamentally depends on accurate keypoint detection and robust descriptor computation, which are essential for tasks such as 3D reconstruction and robot localization. However, performance in low-light conditions remains a significant challenge. To address this, we propose RLFED-NET, a robust network for feature extraction and descriptor computation in low-light scenarios. Firstly, we introduce a differential enhancement network (DENet), which applies differential convolution to feature extraction in matching networks for the first time. Combined with reparameterization techniques, DENet effectively enhances local detail-capturing capability and computational efficiency. Secondly, to tackle the limited descriptor representation in low-light environments, we design the detail information fusion (DIF) module. This module innovatively incorporates multi-scale feature fusion into feature extraction and descriptor computation, preserving fine-grained local details while amplifying high-level semantic features, thereby significantly improving descriptor performance under low-light conditions. Experimental results demonstrate that RLFED-NET outperforms existing methods in homography estimation accuracy (Reprojection error E<3 and E<5) and feature matching performance, exhibiting superior robustness and broader applicability.