The Computer Vision-based Tolerancing Callout Detection Model

Abstract Tolerancing symbols play an important role in mechanical product drawings, and they directly determine the functions, mating properties, interchangeability and working life of geometrical products. A symbolic tolerancing callout containing a set of symbols represents a set of pre-ordered operations with attributes. It includes an amount of knowledge from the drawing and standard documents, which is often reconstructed manually by engineers. Thus, at the same time, a symbolic tolerancing callout makes it difficult for the end-user to understand and interpret these callouts manually. To this end, this study puts forward a tolerancing callout detection model via the use of off-the-shelf costumer-grade cameras on current mobile devices for extracting and recognizing tolerancing callout blocks and symbols in them intelligently. This model has four core components: image preprocessing, callout location and extraction, symbol and character segmentation, and deep learning-based symbol recognition. The image preprocessing component is developed to remove the interferences on the target technical drawings through the corresponding morphological methods. This study proposes a novel solution on callout block locations and extractions in callout intensive scenarios since the callout locations and extractions can directly affect the accuracy of symbol and character recognitions. Then, Huff Transform and improved projection methods have been devised to symbol and character segmentations. Finally, this study constructed a convolutional neural network (CNN) to train a symbol recognition model. The experimental results show that the proposed model gains applicability on intelligent callout extractions and the corresponding symbol recognitions.