基于语义分割和空间上下文信息的防震锤识别

Shock hammer recognition based on semantic segmentation and spatial context information

摘要:

针对目前输电线路中防震锤部件识别精确率低，缺陷无法诊断，未充分利用其空间上下文信息的问题，提出结合DeepLabV3+语义分割网络与防震锤的空间上下文关系对其进行识别与缺陷诊断。利用图像分块和数据集预处理提高DeepLabV3+网络分割精度，将防震锤与其周围部件分割出来后，建立其空间上下文关系缩小防震锤的识别范围，提高其识别精确率。实验结果表明，图像分块与预处理能够将DeepLabV3+网络的分割精度提升到93.4%以上，DeepLabV3+网络可以有效的识别正常防震锤与缺陷防震锤，识别召回率可以达到87%以上，建立防震锤与周围部件的空间上下文关系能够提高其识别精确率到90%以上。

Abstract:

In view of the present shock hammer parts recognition precision rate is low, did not make full use of the relationship of the components of space around problems, put forward to combine the semantic segmentation of segmentation accuracy of the optimal network DeepLabV3+ space context relationship with shock hammer. DeepLabV3+ network segmentation accuracy is improved by image segmentation and data set preprocessing, after the shock hammer and its surrounding parts are separated, the spatial context relationship is established to narrow the identification range of the shock hammer and improve its recognition accuracy. Experimental results show that image segmentation and preprocessing can improve the segmentation accuracy of DeepLabV3+ network to more than 93.4%, DeepLabV3 + network can effectively identify the normal and defective hammer, identify the recall rate can achieve 87% above, establish shock hammer with the surrounding parts of spatial context can improve the identification precision rate to more than 90%.