基于暗光环境的实时目标检测算法研究与应用

Research and Application of Real-Time Object Detection Algorithm Based on Low-Light Environments

摘要: 为了提升视觉任务在复杂暗光环境中的鲁棒性，并满足边缘计算平台对实时目标检测的需求，本文提出了一种改进的轻量化暗光图像处理网络(RPC-YOLO)。该网络前端采用基于Retinexformer改进的RetinexIGTNet对暗光图像进行快速处理，并结合YOLOv5n实现高效的端到端目标检测。此外，模型引入部分卷积(C3-fasterblock)减少计算冗余，跳过无用区域，显著降低参数量并提升推理速度。同时，设计级联注意力机制(CGA)增强特征提取精度和计算效率。实验表明，改进后的模型在公开数据集上的mAP值较原YOLOv5n提升了23%，参数量减少12万，真实场景检测速度(FPS)达到44.65，满足边缘计算平台的实时性需求。该模型为暗光环境下的边缘计算任务(如视觉引导的机械臂抓取)提供了更高效的解决方案。

Abstract: In order to enhance the robustness of visual tasks in complex low-light environments and enable to enhance the robustness of visual tasks in complex low-light environments and meet the real-time object detection requirements of edge computing platforms, this paper proposes an improved lightweight low-light image processing network (RPC-YOLO). The front end of the network uses the RetinexIGTNet, which is based on improvements to the Retinexformer, to quickly process low-light images and combines with YOLOv5n to achieve efficient end-to-end object detection. In addition, the model introduces partial convolution (C3-fasterblock) to reduce computational redundancy by skipping irrelevant regions, significantly reducing the number of parameters and improving inference speed. Meanwhile, a cascaded group attention mechanism (CGA) is designed to enhance the precision and computational efficiency of feature extraction. Experiments show that the improved model has increased the mAP value on public datasets by 23% compared with the original YOLOv5n, reduced the number of parameters by 120,000, and achieved a detection speed (FPS) of 44.65 in real scenes, meeting the real-time requirements of edge computing platforms. This model provides a more efficient solution for edge computing tasks in low-light environments (such as visually guided robotic arm grasping).