基于上下文语义分割信息的缺销螺丝部件识别

Recognition of Defective Screw Parts Based on Context Semantic Segmentation Information

摘要:

针对目前图像视觉领域对输电线路缺销螺丝部件研究较少，且在传统图像处理方法上，螺丝的识别精度不高等问题。文章采用一种基于上下文语义分割信息的缺销螺丝识别方法，在Deeplab v3+网络的基础上，对输电线路数据集进行图像裁剪分块和自适应Gamma校正增强预处理，将缺销螺丝识别的mIoU提升了17%左右；对于普通螺丝误识别，提出了结合上下文语义分割信息的方法，将分割出缺销螺丝区域分别和周围若干部件区域进行拓扑关系分析，根据拓扑关系类别排除误识别到的普通螺丝。通过多组实验结果表明，采用预处理和结合上下文语义信息的缺销螺丝识别方法要优于Deeplab v3+算法。

Abstract:

In view of the current field of image vision, there are few studies on the defective screws in transmission lines, and in the traditional image processing methods, the recognition accuracy of screws is not high. This article adopts a method of identifying defective screws based on contextual semantic segmentation information. On the basis of Deeplab v3 + network, image cropping and blocking and adaptive Gamma correction enhanced preprocessing are performed on the transmission line data set to mIoU for identifying defective screws Increased by about 17%; for misidentification of common screws, a method of combining context semantic segmentation information is proposed, and the area of the missing screw is segmented and several surrounding component areas are analyzed for topological relationship, and the misidentified common screws are excluded according to the type of topological relationship . The results of multiple experiments show that the method of identifying defective screws using preprocessing and combining contextual semantic information is superior to Deeplab v3 + algorithm.