基于改进SOLOv2和虚拟数据增强的输电线路实例分割模型

Instance Segmentation Model for Transmission Line Based on Improved SOLOv2 and Virtual Data Enhancement

摘要:

针对目前将绝缘子自爆缺陷检测和输电线、杆塔检测视作为两个独立的任务，存在对缺陷以及场景理解不充分的问题。提出一种改进的SOLOv2的实例分割方法，可以在航拍图像中同时检视正常绝缘子、自爆缺陷绝缘子、杆塔和输电线。由于绝缘子，杆塔、输电 线之间的长径比相差巨大，该方法使用HRNet替代原始SOLOv2中的ResNet+RPN结构，可以更好地实现多尺度检测，并且在残差结构中引入了可变形卷积DCN，可更好地检测形状细长的绝缘子和输电线。此外，为了提高数据标注工作效率、扩大数据量并节约时间成本，利用虚幻引擎和AirSim自动生成数据集标签的功能，制作了部分虚拟数据集以增强真实数据集。经过实验测试，在构建的数据集中，所提出的方法相较于原始的SOLOv2在平均准确率上提升了8.7%，在平均交并比上提升了8.5%，也优于其他现有实例分割方法。

Abstract:

Currently, the detection of insulator self-explosion defects and the inspection of transmission lines and towers are taken as two separate tasks, resulting in insufficient understanding of defects and scenarios. Aiming at these problems, an improved instance segmentation method based on SOLOv2 is proposed, which allows for the simultaneous inspection of normal insulators, self-explosion defect insulators, towers and transmission lines in aerial images. Due to the significant aspect ratio differences among insulators, towers and transmission lines, this method replaces the ResNet+RPN structure in original SOLOv2 with HRNet, which better achieves multi-scale detection. Additionally, deformable convolution is introduced into the residual structure to better detect the elongated shapes of insulators and transmission lines. Furthermore, in order to enhance data annotation efficiency, increase data volume and save time, a portion of the virtual dataset is created using unreal engine and AirSim to augment the real dataset. Experimental tests show that, in the constructed dataset, the proposed method improves the average accuracy by 8.7% and the average intersection over union (IoU) by 8.5% compared to the original SOLOv2, and it also outperforms other existing instance segmentation methods.