面向嵌入式设备部署的轻量化织物瑕疵检测算法

Lightweight fabric defect detection algorithm for embedded device deployment

针对现有织物瑕疵检测算法参数量大、计算复杂度高,难以部署在计算资源有限的嵌入式设备上的问题,提出一种基于YOLOv5s改进的轻量化织物瑕疵检测算法SSPY(ShuffleNetv2-S,SimAM-S,Pruning-P,YOLOv5s-Y)。首先,主干特征提取网络采用ShuffleNetv2网络,实现模型的轻量化。在主干网络和小目标检测层引入SimAM无参注意力机制,在不增加额外参数量的情况下增强算法的特征提取能力。通过结合稀疏训练评价特征提取层中卷积核的重要性进行剪枝的方法,进一步实现模型压缩。最后,将SSPY算法部署到瑞芯微RK3568平台上,完成织物瑕疵实时检测算法在嵌入式设备上的部署。在织物瑕疵数据集上进行多组对比实验。实验结果表明,SSPY与YOLOv5s相比,平均精度均值mAP值提升了0.8%,参数量下降了80.3%。将SSPY部署在RK3568上,检测速度可达49 FPS,满足了织物瑕疵检测算法在工业应用中实时性、嵌入式设备部署等需求。

In response to the issues of large parameter quantity,high computational complexity,and difficulty in deploying the existing fabric defect detection algorithms on embedded devices with limited computing resources,an improved lightweight fabric defect detection algorithm SSPY based on YOLOv5s was proposed.Firstly,the ShuffleNetv2 network was used as the backbone feature extraction network to achieve lightweight model.The SimAM no-parameter attention mechanism was introduced in the backbone network and small target detection layer,enhancing the feature extraction capabilities of the algorithm without adding additional parameters.Model compression was further achieved by pruning based on the importance of the convolution kernel in the sparse training evaluation feature extraction layer.Finally,the SSPY algorithm was deployed on Rockchip RK3568 platform,and the deployment of the fabric defect real-time detection algorithm was completed on embedded devices.Multiple comparison experiments were carried out on fabric defect data set.Experimental results show that compared with YOLOv5s,SSPY′s mAP increases by 0.8%and the number of parameters decreases by 80.3%.When SSPY was deployed on RK3568,the running speed can reach 49 FPS,which met the needs of real-time performance and embedded device deployment of fabric defect detection algorithms in industrial applications.