부산대학교 도서관

With the development of smart grids, the methods of using cameras or robots for image acquisition and recognition of meters are gradually replacing traditional manual inspections. To enable more efficient meter recognition by deep learning models, the quality of training data is of paramount importance, with data imbalance being a major challenge in rough data. We propose a data augmentation method for meter instance segmentation under a data balancing strategy, called meter balanced object paste (MeterBOP), which pastes meter objects into images of different scenes according to predefined rules, thereby optimizing the data distribution. The approach selects the pasting meter types by class-balanced sampling and scales them to various sizes for scale balance. The pasting regions are generated according to the distribution of existing objects in the image, thus achieving spatial balance while avoiding severe occlusion caused by introducing new objects. Experiments demonstrate that our method improves the performance of Mask R-CNN by 3.1 Box mAP and 1.0 Mask mAP on power meter image-mask (PMI-M) dataset. Furthermore, the method can provide stable performance improvements under various detectors and conditions, exhibiting superior performance enhancement compared to other data augmentation methods.