基于轻量化YOLOv7-DenseBlock的绝缘子检测算法

An Insulator Detection Algorithm Based on Lightweight YOLOv7-DenseBlock

中文关键词:绝缘子检测  YOLOv7  二阶拉普拉斯算子  DenseBlock  FReLU

keywords:insulator detection  YOLOv7  second order Laplacian  DenseBlock  FReLU

中文摘要:绝缘子巡检过程中要求快速、准确地检测出绝缘子及其缺陷，当前绝缘子检测算法存在检测速度慢、精确度低的问题。提出一种基于轻量化YOLOv7-DenseBlock的绝缘子检测算法。将原图通过二阶拉普拉斯算子进行锐化处理，以加强原图绝缘子的特征；在YOLOv7主干网络上加入DenseBlock模块，以提高梯度的反向传播，并增强特征重用的效率；将激活函数替换为FReLU，以解决PReLU、LeakyReLU激活函数对空间信息不敏感的问题，提高模型鲁棒性，降低模型的错检率、漏检率。实验结果表明，改进的模型更具轻量化，检测速度更快：相较于YOLOv7，其参数量减少了7.38%，低至33.79M；检测速度提升了9.55%，达到45.45 FPS。与YOLOv5-CBMA相比，其精度提升了1.77%。

Abstract:The insulator inspection process requires rapid and accurate detection of insulators and their defects， but current insulator detection algorithms suffer from slow detection speed and low accuracy. This paper proposes an insulator detection algorithm based on lightweight YOLOv7-DenseBlock. Firstly， the original image is sharpened using a second-order Laplacian operator to enhance the features of insulators in the original image. A DenseBlock module is incorporated into the YOLOv7 backbone network to improve gradient backpropagation and enhance feature reuse efficiency. The activation function is replaced with FReLU to address the insensitivity of PReLU and LeakyReLU activation functions to spatial information， thereby improving model robustness and reducing false detection and missed detection rates. Experimental results demonstrate that the improved model is more lightweight and achieves faster detection speed: compared to YOLOv7， its parameter count is reduced by 7.38% to 33.79M， and the detection speed is increased by 9.55% to 45.45 FPS. Compared to YOLOv5-CBMA， its accuracy is improved by 1.77%.