基于改进YOLOv7-tiny的多种类绝缘子检测算法

Multi-type insulator detection algorithm based on improved YOLOv7-tiny

针对现有绝缘子检测算法识别种类单一、定位精度差、鲁棒性差等问题,提出了一种改进YOLOv7-tiny的多种类绝缘子检测算法。首先,使用K-means++算法对先验框进行重聚类,获得更适用于多种类绝缘子数据集的先验框尺寸;其次,采用了基于动态非单调的聚焦机制设计的WIoUv3损失函数,解决训练过程中正负样本不均衡问题。在网络结构上,首先在骨干网络使用跨阶段特征融合模块(CFFCB)捕获更多的上下文信息,对一些受到遮挡的绝缘子实现精准检测;同时,在颈部网络,提出了空间金字塔池化模块SPPCSPF替换了原有的SPPCSP,有效提高绝缘子与背景接近时的检测成功率,有效改善漏检情况。经过实验测试,与YOLOv7-tiny相比,改进后的网络模型的mAP提高了2.1%,达到了97.6%,有效提高了多种类绝缘子的检测精度。最后,利用改进后算法的检测结果在UR5机械臂上进行了抓取实验,实际抓取的成功率在90%左右,验证了算法的可行性。

Aiming at the problems of limited insulator type recognition,poor positioning accuracy and lack of robustness in existing insulator detection algorithms,a multi-type insulator detection algorithm based on improved YOLOv7-tiny is proposed.Firstly,the K-means++algorithm is used to recluster the anchor box to obtain the anchor box size which is more suitable for multi-type insulator datasets.Secondly,the WIoUv3 loss function based on the dynamic non-monotone focusing mechanism is designed to address the imbalance between positive and negative samples in the training process.On the network structure,firstly,the Cross-stage Feature Fusion-ConvNeXt Block(CFFCB)is used to capture more context information at the Backbone,and some occluded insulators are accurately detected.At the same time,at the Neck,the SPPCSPF(Spatial Pyramid Pooling Cross Stage Partial-Fast)is proposed to replace the original SPPCSP,(Spatial Pyramid Pooling Cross Stage Partial),which effectively improves the detection success rate when the insulator is close to the background,and effectively improves the missed detection situation.After experimental testing,compared with YOLOv7-tiny,the mAP of the improved network model is increased by 2.1%,reaching 97.6%,which effectively improves the detection accuracy of various insulator types.Finally,the grabbing experiment is carried out on the UR5 manipulator by using the detection results of the improved algorithm.The actual grabbing success rate is about 90%,which verifies the feasibility of the algorithm.