关键词: 水工混凝土材料, 文本检测, 文本识别, 深度学习, 领域知识

Abstract: During the construction of a hydropower project, a large number of non-editable documents for hydraulic concrete materials are generated. Using manual interpretation methods to obtain texts is time-consuming, laborious and accuracy-uncontrollable, making it difficult to meet the demand for information management of material data. This paper develops an intelligent interpretation method for non-editable texts of hydraulic concrete materials. First, we construct a text detection model, HC-PSENet, based on pixel level segmentation, which integrates the backbone network of PP-HGNet to achieve accurate detection of text lines. Then, a professional corpus is created based on the domain knowledge to realize accurate character mapping. We construct a text recognition model HC-CRNN for hydraulic concrete materials, using detection text boxes and the professional corpus as its inputs, and adopt the backbone network of ResNet and the improved loss function C-CTC Loss to improve the accuracy of character classification. Finally, a transfer learning strategy is adopted to train the model with the self-designed dataset as an example; the effectiveness and superiority of our new method is verified through ablation and comparative experiments. The results show that it has a harmonic mean of 0.985 for detecting text regions and its accuracy of text recognition reaches 90.62%. It has an overall performance superior to classical methods and would provide new technical means for the automated reuse of non-editable text resources in concrete materials.

Key words: hydraulic concrete materials, text detection, text recognition, deep learning, domain knowledge