An Improved Cascade R-CNN Method for Meter Box Defect Detection

In order to address the challenges associated with defect detection in meter boxes, this research presents an improved Cascade R-CNN algorithm for effective defect identification. Initially, various attention mechanisms are evaluated within the context of meter box defect detection using the Cascade R-CNN framework. Through comprehensive experimental comparisons, the attention mechanism exhibiting superior detection accuracy is selected and integrated into the backbone network as an attention module, enabling better handling of defects with varying scales in meter boxes. Subsequently, field detection and application verification are conducted in a power distribution area of a province to validate the effectiveness of the enhanced Cascade R-CNN algorithm for meter box defect detection tasks. The experimental results demonstrate that the proposed method significantly enhances the efficiency of defect detection work by enabling detection personnel to rapidly screen and locate defective meter boxes using handheld terminals under practical conditions. Compared with the ResNet-50-based method, our structure improves mAP by 1.3 without increasing the number of parameters and GFlops.