Deep Learning Based Character Recognition for Digital Meters

To address the challenges of suboptimal accuracy and latency in digital meter character recognition within industrial environments, this paper proposes an enhanced optical character recognition (OCR) method that improves both recognition accuracy and computational efficiency. The proposed approach comprises two main components: (1) architectural optimization of the PaddlePaddle OCR (PPOCR) system through a lightweight redesign, where the original region-based convolutional neural network (RCNN) is replaced by a Transformer architecture incorporating self-attention mechanisms within a MobileNet-based framework-forming a hybrid network with dynamic feature extraction capabilities; and (2) systematic dataset preparation, including manual annotation and adaptive data augmentation, to ensure effective training. Experimental results show that the optimized model achieves a mean processing time of 70 ms per image (a 50% reduction compared to the original) on an NVIDIA T4 platform, while attaining over 95% character recognition accuracy (an absolute improvement of 10% over the baseline), demonstrating its effectiveness for industrial digital meter recognition tasks.