YOLO-RSLW: Research on Optimization of Lightweight YOLO11 Object Detection Algorithm

Abstract

To address the challenges of excessive parameters and high computational complexity in YOLO-series models under resource-constrained environments, this paper proposes an improved YOLO11 framework named YOLO-RSLW, which focuses on the synergistic optimization of lightweight design and accuracy enhancement. The proposed approach systematically reconstructs the model architecture by targeting two critical components: feature fusion and detection head design, thereby significantly reducing model complexity while improving detection performance. Specifically, a lightweight RepNCSPELAN4Lighter module is designed based on reparameterization techniques] to replace the original C3k2 modules in both the backbone and neck networks. The SPPF and C2PSA modules are removed, and an SPPELAN module is introduced to enhance feature representation capability. An efficient LiteShiftHead is proposed to optimize the detection pipeline through task-decoupled group-wise processing and lightweight convolutions. Furthermore, the WIoU loss function is adopted to replace the original loss computation, enhancing the accuracy and robustness of bounding box regression. Experimental results on the DOTAv1.5 aerial image dataset demonstrate that the improved model reduces the parameter count by 37.3% (down to 5.92M), while achieving a 1.1 percentage point improvement in both mAP50 and mAP50-95 (with mAP50 reaching 38.2%), realizing a synergistic optimization of accuracy and efficiency. Ablation studies validate the effectiveness of each proposed module, among which RepNCSPELAN4Lighter and LiteShiftHead contribute most significantly to model compression, while the WIoU loss further improves localization accuracy. This study provides a more efficient detection solution for embedded platforms and real-time vision applications.