基于YOLOv4的绝缘子检测算法

YOLOv4-Based Algorithm for Insulator Detection

针对绝缘子检测过程中图像易出现失真,以及模型受外界环境影响较大的问题,提出一种基于YOLOv4的绝缘子检测算法。该算法首先采用自适应伽马变换自动调整绝缘子图像的明亮度,然后使用YOLOv4网络学习绝缘子不同层次的特征表示。考虑到YOLOv4网络需要输入固定的图像大小,而强行对图像进行拉伸会使得目标出现扭曲,故采用可变形卷积替换传统卷积的方法,从而提高模型的特征提取能力,最终输出绝缘子的位置信息及其类别。在中国电力绝缘子公开数据库中进行仿真实验,结果表明该算法的测试精度和检测速度分别达到了93.2%和43FPS。该算法的总体性能优于Faster RCNN、YOLOv3、CornerNet等常用算法。

Aiming at the problem that the image is easy to distort in the process of insulator detection and the model is greatly affected by the external environment,an insulator detection algorithm based on YOLOv4 is proposed.Firstly,adaptive gamma transform is used to automatically adjust the brightness of the insulator image.Then,the YOLOv4 network is used to learn the feature representation at different levels of the insulators.Considering that the YOLOv4 network needs to input a fixed image size,and the forced stretching of the image will distort the target,deformable convolution is adopted to replace the traditional convolution method,so as to improve the feature extraction ability of the model.Finally,the final experimental results are obtained after the insulator position information and its categories are output.The experimental simulation on the public database of power insulators in China shows that the test accuracy and detection speed of the proposed algorithm are 93.2%and 43FPS,respectively.The overall performance of the algorithm is better than Faster RCNN,YOLOv3,CornerNet and other common algorithms.