摘要
针对在复杂边缘计算场景下外力破坏目标检测识别精度低与实时性差的问题,提出YOLO-ERFA轻量化目标检测算法。该算法采用跨阶段残差结构的CSPDarkNet53-Tiny作为特征提取网络,在保证模型轻量化的同时提高检测准确率;在此基础上,通过改进空间金字塔池化并融合高效通道注意力机制构建特征增强层以提升模型精度和对多尺度目标的检测能力,并在训练阶段使用改进的Mosaic算法增加样本背景虚化以提高模型在复杂场景下的抗干扰能力。实验结果表明,该方法在测试集上检测平均准确率达到了91.58%,在Jetson TX2平台推理速度达30 FPS,且模型内存大小仅为26.50 MB,提高了算法在边缘计算设备上部署的可行性。
The lightweight target detection algorithm YOLO-ERFA was proposed to solve the problem of low accuracy and poor real-time performance of target detection for heavy machinery in complex edge computing scene.CSPDarkNet53-Tiny network with cross-stage residual structural feature extraction was used for improving the detection accuracy of the lightweight model.On this basis,the feature enhancement layer was constructed by improving the spatial pyramid pooling algorithm and integrating the efficient channel attention mechanism to improve the accuracy of the model and the detection ability of multi-scale targets.In the training stage,the improved Mosaic algorithm was used to improve the anti-interference ability of the model in complex scenes with increasing the sample background virtualization.The experimental results on the test data set show that the average accuracy of the algorithm in this paper reaches 91.58%,the computing speed reaches 30 FPS on the Jetson TX2 platform,and the memory occupation of the network model is only 26.50 MB,which improves the feasibility of deploying the algorithm on the edge computing device.
作者
蓝向州
卢泉
陈桥
LAN Xiang-zhou;LU Quan;CHEN Qiao(School of Electrical Engineering,Guangxi University,Nanning 530004,China)
出处
《广西大学学报（自然科学版）》
CAS
北大核心
2022年第5期1363-1373,共11页
Journal of Guangxi University（Natural Science Edition）
基金
国家自然科学基金项目(61863002)。
关键词
目标检测
边缘计算
轻量化网络
注意力机制
target detection
edge computing
lightweight network
attention mechanism
