Robust Pointer Meter Reading Recognition Method Under Image Corruption

With the development of smart grids, vision-based meter reading recognition methods are gradually replacing traditional manual inspections. In substations, meter reading faces image corruption challenges caused by various factors such as strong electromagnetic interference, lighting variations and weather conditions. The existing methods show significant performance degradation in noisy environments. To address this, we propose a robust pointer meter reading recognition method under image corruption. We first design image corruption augmentation (ICA), which significantly enhances the resilience of the model to disturbances. We introduce mask scoring convolution region-based convolutional neural network (MSC R-CNN) to segment pointer and scale masks on dial. MSC head improves localization and segmentation accuracy, while the balanced aggregation feature pyramid (BAFP) fuses features and enables multi-scale predictions. The inclusion of the global context (GC) block mitigates the impact of interference. For meter readings, we develop scale area proportion (SAP) reading method to process pointer and scale masks. The experimental results demonstrate that MSC R-CNN achieves 64.2 Mask mAP and 58.4 Mask mPC, surpassing the 58.4 Mask mAP and 26.0 Mask mPC of Mask R-CNN. SAP reading method achieves a successful meter reading rate (SMR rate) of 94.6% and maintains an average SMR rate of 92.4% under image corruption. The proposed method manifests robust meter reading recognition under image corruption. It can contribute to the further upgrading of smart grids. The code is available at https://github.com/ZhaolinWang0110/MeterRecognition.