基于YOLOv8和DeepLabv3+的指针仪表读数识别  

Recognition of Point Meter Readings Based on YOLOv8 and DeepLabv3+

作　　者：吴肖 赵洪泉 杨鹏 张景元 石波 Wu Xiao;Zhao Hongquan;Yang Peng;Zhang Jingyuan;Shi Bo(Huaneng Nuclear Energy Technology Research Institute Co.,Ltd.,Shanghai 200126,China;Shouan Industrial Fire Protection Co.,Ltd.,Beijing 101300,China;Beijing Xiangcheng Technology Development Co.,Ltd.,Beijing 101300,China)

机构地区：[1]华能核能技术研究院有限公司,上海200126 [2]首安工业消防有限公司,北京101300 [3]北京享成技术发展有限公司,北京101300

出　　处：《机电工程技术》2024年第6期240-244,共5页Mechanical & Electrical Engineering Technology

基　　金：华能集团总部科技项目(HNKJ21-H22)。

摘　　要：指针仪表作为一种重要的检测工具在核电领域中被广泛使用,针对不同类型、不同距离条件下指针仪表读数识别准确性低和检测速率低下的问题,提出了一种基于YOLOv8和DeepLabv3+的仪表读数识别方法。为了提高DeepLabv3+模型的图像输入质量,选择了推理速度快且准确的YOLOv8检测器定位仪表区域并裁剪后作为输入图像用于仪表识别。针对仪表指针识别准确率低和检测速度慢的问题,在DeepLabv3+模型的基础上,将骨干网络替换为MobileNetv3,并设计ECA+模块将其SE模块替换,降低模型参数的同时提高识别精度;将解码器的4倍上采样替换为两个2倍上采样,浅层特征图与编码器相应尺寸的特征进行拼接融合,并引入CBAM模块,提高指针分割的准确性。试验阶段,采用自制核电厂指针式仪表图像数据集,实验结果表明,该方法在仪表精度为2.5的级别下达到了94%的识别准确率,平均误差为1.542%,平均总耗时为0.57 s,具有较好的性能表现。Pointer instrument is widely used in the field of nuclear power as an important detection tool.In view of the low accuracy and detection rate of pointer instrument reading recognition under different types and distances,a method for instrument reading recognition based on YOLOv8 and DeepLabv3+is proposed.In order to improve the image input quality of the DeepLabv3+model,the YOLOv8 detector with fast reasoning speed and accurate positioning is selected to locate the instrument area and crop it as the input image for instrument recognition.In view of the low accuracy and slow detection rate of instrument pointer recognition,on the basis of the DeepLabv3+model,the backbone network is replaced with MobileNetv3,and the ECA+module is designed to replace its SE module,reducing model parameters while improving recognition accuracy.The quadruple upsampling of the decoder is replaced with two double upsampling,and the shallow feature map is spliced with the corresponding size feature of the encoder.The CBAM module is introduced to improve the accuracy of pointer segmentation.In the experimental stage,using self-made image dataset of nuclear power plant pointer instrument,the experimental results show that the method achieves a recognition accuracy of 94%at an instrument accuracy level of 2.5,with an average error of 1.542%and an average total time consumption of 0.57s,which has good performance.

关 键 词：核电 指针式仪表 YOLOv8 DeepLabv3+ 读数识别 

分 类 号：TH7[机械工程—仪器科学与技术]