3 Method
Our proposed framework intends to locate insulator areas, determine if
these insulators are faulty, and identify which categories these
insulators belong to. It needs to be emphasized that the framework
explores the feasibility of this task under the conditions that a
portion of annotations are inaccessible and sample sizes of various
classes are imbalanced. In detail, our framework is based on the Faster
RCNN pipeline [27], illustrated in Figure 1. We improved the Faster
RCNN by combining PU learning with Region Proposal Network (RPN) and
incorporating focal loss into the Region Of Interest (ROI) Head.