Progressive Feature Fusion and Refinement Network for Substation Rotating Object Detection

Abstract

Real-time substation object detection is of great significance to ensuring the safe and stable operations of the power grid. Considering that the substations are complex in the background and the targets are distinct in sizes, shapes and rotation angles, we propose a progressive feature fusion and refinement network (PF2RNet) for substation rotating object detection. In the network, ResNeSt50 is used as the backbone to improve the feature extraction ability, and the deconvolution feature fusion module is designed to generate richer semantic information. To perform better in substation scenes, the rotating anchors are used to reduce the Intersection over Union between anchor boxes. Besides, the feature refinement module is introduced to realize the regression process from coarse to fine, strengthen the feature information of the object location, and then alleviate the feature misalignment. Finally, experiments demonstrate that the mAP of PF2RNet on the substation multi-object dataset reaches 89.3%, which is improved by 5.2% compared to RetinaNet.