Abstract: Label detection of electrical cabinets in the distribution room of chemical plants is an important part of intelligent and autonomous inspection. The traditional manual detection method has the disadvantages of low detection efficiency, high rate of missed detection and high false detection rate, and risky for workers. The OCR-based deep learning model can efficiently detect the position of the text box of the label on the electrical cabinet and accurately identify the text content. On the self-built electrical cabinet dataset, a text detection model DBNet and a text recognition model ABINet are trained to detect the position of the electrical cabinet text box. The accuracy of the DBNet model in detecting text boxes on the test set can reach 96%, the accuracy of ABINet in recognizing text can reach 89%, and the average detection speed of the model can reach 1 image per second under the RTX2080Ti graphics card. The methods applied in the experiment meet the requirements of industrial detection accuracy and detection speed.

Key words: Optical Character Recognition, Text Detection, Text Recognition, Deep Learning

摘要： 化工厂配电室电柜标签检测是智能智能自主巡检中的重要一环。传统人工检测方法存在检测效率低、漏检错检率高、危险性高等缺点。基于OCR的深度学习模型可以高效的检测出电柜上标签的文本框位置并准确的识别出文本内容。在自建电柜数据集上,训练一个可以检测出电柜文本框位置的文本检测模型DBNet和一个文本识别模型ABINet。DBNet模型在测试集上检测文本框的准确率可达96%,ABINet识别文本的准确率可达89%,在RTX2080Ti显卡下,模型的平均检测速度可达1张/s。实验中应用的方法满足工业上的检测精度与检测速度需求。

关键词: 光学字符识别, 文本检测, 文本识别, 深度学习