An Improved YOLOv7-Based Multi-Strategy Model for Low-Light Object Detection

Currently, the field of low-light object detection faces multiple challenges in complex road environments, including dense object distributions, diverse scales, and inadequate illumination, which results in unsatisfactory performance of general-purpose object detection algorithms in real-world applications. To address these issues, this work proposes a multistrategy low-light detection model named Multi-Strategy LowLight Detection YOLO (MSLD-YOLO) based on the YOLOv7 framework. The proposed model introduces several key innovations: First, a Recursive Gated Convolution Block (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{g}^{\mathrm{n}}$</tex> Block), is incorporated into the backbone network to enhance the conventional self-attention mechanism's capacity in modeling second-order spatial information. Second, a Normalization-based Attention Module (NAM) is embedded within the feature reuse layer to achieve a more effective fusion of shallow and deep feature information. Third, a Dual-Branch Decoupled Detection Head is utilized to separately predict the class labels, bounding boxes, and objectness scores, which effectively accelerates model convergence and improves detection accuracy. Finally, a novel localization regression loss function that comprehensively accounts for multiple factors is adopted, further enhancing the detection performance without increasing additional model parameters. Experimental results on the BDD100K dataset demonstrates that the improved model achieves improvements of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2. 2 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 6 \%}$</tex> in [email protected] and [email protected]:0.95, respectively, while significantly reducing both false positive and false negative rates under lowlight conditions.