基于YOLOv5的绝缘子图像识别算法轻量化改进研究

Lightweight Improvement of Insulator Image Recognition Algorithm Based on YOLOv5

为解决传统神经网络图像识别模型无法在无人机等微型嵌入式设备有效部署的问题,基于YOLOv5算法提出一种针对无人机嵌入式平台的电力巡检绝缘子图像轻量化识别算法。以ShuffleNetV2为骨干网络,并使用深度可分离卷积替换传统卷积减少参数量,降低算法对物理内存的需求。同时针对绝缘子特征对图像预处理部分进行改进,使用PeleeNet网络中的Stem结构代替Focus结构,加快训练速度,减轻嵌入式平台CPU的计算压力。另外,在残差网络中引入Transformer注意力机制,提高算法对遮挡目标的提取能力。所提算法能够在支持Arm架构的树莓派4B+平台部署,检测准确率可达91.6%,对遮挡绝缘子检出率能达到96%。处理尺寸为640×640的单张图片平均检测时间0.97 s。

In order to solve the problem that the traditional neural network image recognition model cannot be effectively deployed in micro-embedded devices such as UAVs,a lightweight recognition algorithm for power inspection insulator images for UAV embedded platforms is proposed based on the YOLOv5 algorithm.The ShuffleNetV2 is used as the backbone network,and depthwise separable convolution is used to replace the traditional convolution to reduce the amount of parameters and reduce the physical memory requirement of the algorithm.At the same time,the image preprocessing part is improved for the characteristics of insulators,and the Stem structure in the PeleeNet network is used instead of the Focus structure to speed up the training speed and reduce the computational pressure of the embedded platform CPU.In addition,the transformer attention mechanism is introduced into the residual network to improve the algorithm's ability to extract occlusion targets.The proposed algorithm can be deployed on the Raspberry Pi 4B+platform that supports the Arm architecture,and the detection accuracy can reach 91.6%,and the detection rate of shading insulators can reach 96%.The average detection time for a single image with a size of 640×640 is 0.97s.