Dark Object Detection Based on Detail Feature Enhancement and Cross-aligned Attention

Images captured in dark and low-light environments present significant challenges for feature extraction algorithms in computer vision, particularly impacting the performance of object detection tasks due to inadequate lighting and severe noise. Previous multi-level feature extraction methods aimed to enhance performance by directly fusing features from different layers but often overlooked the misalignment and degradation of cross-scale features during fusion. We hypothesize that reducing differences between cross-level features before fusion can yield stronger feature representations. In this paper, we propose a Detail Feature Enhancement Block (DFEB) and a Detail Feature Enhancement Network (DFEN) based on it, specifically designed to enhance low-texture detail features in low-light images, thereby improving the model’s ability to capture image features from the source. Additionally, we introduce a Cross-Aligned Attention-based (CAFF) multi-level feature fusion method. Specifically, we compute cross-attention between adjacent feature layers before fusion and use deformable convolutions to integrate degraded differential feature maps into the neighboring intermediate image features, eliminating differences between multi-level detail features. This ensures that subsequent feature fusion captures more comprehensive and accurate low-light image information. By combining these components, we present a new framework for dark object detection based on detail feature enhancement and cross-aligned attention. Extensive experiments on the ExDark dataset demonstrate the effectiveness of our method in enhancing feature extraction, yielding state-of-the-art results in dark object detection tasks.