A Weakly Supervised Instance Segmentation Approach for Insulator Thermal Images Incorporating Sparse Prior Knowledge

The efficiency of substation inspection can be enhanced by employing an instance segmentation method to preprocess the substation thermal images. Nevertheless, the current task of thermal image segmentation for substation insulators faces difficulties in annotation and detection. To address these challenges, a weakly supervised instance segmentation method is proposed. Firstly, a general sparse prior knowledge is introduced to enhance the detection accuracy on insulators thermal image datasets without incurring additional costs. Then, the model incorporates an improved backbone with pyramid structure, and replaces the original feature pyramid network with a bidirectional feature pyramid network. Finally, to verify the application of the improved model in the electric power visual scene, an automatic annotation framework based on predicted mask is proposed. The pixel-level annotation is transformed into the box-level annotation, thereby addressing the issue of complex and time-consuming annotation of segmentation data in the electric power domain. The test results demonstrate that the improved weakly supervised instance segmentation method has achieved significant improvement in accuracy on both self-built insulators thermal image datasets and two publicly available datasets. Moreover, the annotated data obtained by the automatic annotation framework has achieved competitive detection accuracy on multiple instance segmentation models.