基于轻量化深度学习网络的数字仪表识别

数字式仪表常用于变电站、工厂等生产环境，是一种直观的设备监测仪器。然而当前数字式仪表的读取方式还依赖于人工巡检，手动记录等，这些传统的巡检方式来监测设备的运行状态大大降低了巡检效率。为了实现传统行业的数字化转型，本文提出基于轻量化深度学习的数字仪表识别方法，通过改进的YOLOv5的目标检测框架，针对数字仪表目标区域在整张图片大小不一致的情况，提出对于感兴趣区域（ROI）的迭代目标检测方法，首次检测将感兴趣区域进行检测并切割统一到相同的尺度，随后迭代检测网络针对感兴趣区域内的字符进行检测并分类，以达到精确读数的目的。为提升多尺度检测性能，本文采用Res2Net模块主干网络中的的残差模块。采用GIoU取代通用的IoU作为位置损失函数加速模型训练效果的收敛。实验表明，改进后的框架实现了99.62%的准确率和99.55%的召回率，相比基线网络分别提升了12.72%和5.85%。通过将框架在边缘计算平台上的终端部署，在实际生产中取代了人工巡检，实现了商业化运行。

Digital Instrument Recognition Based on Lightweight Deep Learning Network

Digital instruments are commonly used in substations, factories, and other production environments as intuitive equipment monitoring instruments. However, the current reading method of digital instruments still relies on manual inspection, manual recording, etc. These traditional inspection methods to monitor the operation status of the equipment greatly reduce inspection efficiency. To realize the digital transformation of traditional industries, a digital instrument recognition method based on lightweight deep learning was proposed in this paper. Through the improved target detection framework of YOLOv5, because of the inconsistent size of the target area of the digital instrument in the whole picture, a new method was proposed for the region of interest ( ROI) iterative target detection method, the first detection detected and cut the region of interest to the same scale, and then the iterative detection network detected and classified the characters in the region of interest to achieve the purpose of accurate reading. To improve the multi-scale detection performance, the residual module in the Res2Net module backbone network was used in this paper. GIoU was used instead of the generic IoU as the position loss function to accelerate the convergence of the model training effect. The experiments show that the improved framework achieves 99.62% accuracy and 99.55% recall, which are 12.72% and 5.85% improvements compared to the baseline network, respectively. Through the terminal deployment of the framework on the edge computing platform, manual inspection is replaced in actual production, and commercial operation is realized.