| 引用本文: | 尹向雷,解永芳,屈少鹏,等.基于动态蛇形卷积和非跨步卷积的绝缘子缺陷检测[J].电力系统保护与控制,2024,52(20):177-187.[点击复制] |
| YIN Xianglei,XIE Yongfang,QU Shaopeng,et al.Insulator defect detection based on dynamic snake convolution and space-to-depth convolution[J].Power System Protection and Control,2024,52(20):177-187[点击复制] |
|
| 摘要: |
| 针对复杂场景下绝缘子缺陷检测存在小目标识别困难的问题，提出基于动态蛇形卷积和非跨步卷积的绝缘子缺陷检测方法。首先，算法引入动态蛇形卷积，构造出符合绝缘子特点的特征提取模块，提高对绝缘子及其缺陷的特征提取能力。然后，采用“空间-深度”的非跨步卷积，减少融合过程中的特征丢失。最后，为进一步降低模型复杂度，对模型进行通道剪枝，减少冗余部分。在绝缘子缺陷数据集上进行实验对比，与基准算法相比，绝缘子的破损、污闪以及自爆缺陷的识别率分别提升了5.7%、2.4%和0.8%，改进算法在绝缘子的检测率上提升了0.5%。同时平均精度均值较改进前提升了2.3%，模型大小降低了50.07%。实验结果表明，改进算法在提高绝缘子缺陷小目标检测精度的同时，有效降低了模型大小，对绝缘子缺陷检测的研究具有一定的参考和应用价值。 |
| 关键词: 绝缘子 缺陷检测 YOLOv7 动态蛇形卷积 “空间-深度”的非跨步卷积 |
| DOI：10.19783/j.cnki.pspc.240583 |
| 投稿时间：2024-05-15修订日期：2024-07-08 |
| 基金项目:国家自然科学基金面上项目资助(62176146)；陕西省教育厅重点科学研究计划项目资助(20JS018)；陕西理工大学人才启动专项资助(SLGRCQD2114) |
|
| Insulator defect detection based on dynamic snake convolution and space-to-depth convolution |
| YIN Xianglei,XIE Yongfang,QU Shaopeng,SU Ni |
| (School of Electrical Engineering, Shaanxi University of Technology, Hanzhong 723000, China) |
| Abstract: |
| There is a problem that small object detection of insulators defects is difficult because of complex scenes. Thus this paper proposes a detection model of insulator defects based on dynamic snake convolution and space-to-depth convolution. First, the algorithm imports dynamic snake convolution to shape the feature extraction module that conforms to the characteristics of insulators, improving the ability of feature extraction for insulators and their defects. Next, this algorithm takes the “space-to-depth” convolution to reduce the features missing during the fusion. Finally, the algorithm is pruned to decrease redundancy in order to further simplify the complexity of model. With experiments on an insulator defects dataset, the recognition rates of broken, pollution-flashover and self-exposing insulators are increased by 5.7%, 2.4% and 0.8% respectively. In addition, the detection rate of insulators is raised by 0.5% contrasted to baseline. Meanwhile, mean average precision is increased by 2.3% and model size is diminished by 50.07%. The results of experiments show that the algorithm can not only promote detection accuracy of a small object insulator defect, but also it can sink model size, which means that the algorithm possesses a certain reference and application value for a survey of insulator defect detection. |
| Key words: insulators defects detection YOLOv7 dynamic snake convolution “space-to-depth” convolution |
