【摘要】 针对深度学习算法难以部署到无人机硬件和小目标易漏检误检问题，提出改进YOLOv8的轻量化小目标识别算法DLS-YOLO。采用深度可分离卷积替代普通卷积轻量化网络，设计多分支空洞卷积结构IRFB增大感受野，在主干网络中加入ECA注意力机制增强特征表达能力，增设160×160尺寸检测头，提高小目标检测能力。在VisDrone2019数据集上，DLS-YOLO比YOLOv8s在mAP@0.5上提高了2.3%，计算量降低11.2GFLOPs，模型尺寸缩减至5.7 M，性能优于其他主流检测方法。在Jetson Xavier NX平台验证算法的有效性，平均单帧处理时间34.4 ms，实现近实时检测。更多还原

【Abstract】 In view of the problems that deep learning algorithm is difficult to be deployed to drone hardware and small targets are easy to be missed and mis-detected, a lightweight small target recognition algorithm, DLS-YOLO, is proposed to improve YOLOv8. This paper adopts the deep separable convolution to replace the ordinary convolution lightweight network, designs the multi-branch cavity convolution structure, IRFB, to increase the receptive field, introduces the ECA attention mechanism to the main network to enhance the ability of feature expression and add the 160×160 detection head to improve the ability of detection of the small targets. On VisDrone2019 data set, DLS-YOLO is 2.3% higher than YOLOv8s in mAP@0.5, the computational load is reduced by 11.2GFLOPs and the model size is reduced to 5.7 M, which is superior to other mainstream detection methods. The effectiveness of the algorithm is verified on the Jetson Xavier NX platform with an average processing time for a single frame of 34.4ms, thus can realize near real-time detection.更多还原