End-to-end Character Sequence Recognition System For Engineering Drawings Captured By Mobile Terminals

In recent years,with the development of information science and technology,the improvement of computer efficiency,and the popularization of the mobile terminal,the application of traditional engineering in the mobile terminal has become more and more widespread.As one of the important elements in the fields of industry and architecture,the intelligent identification and matching of engineering drawing have great research value,receiving extensive attention from domestic and foreign engineering researchers and experts.It has achieved outstanding results in intelligent recognition of complex scenes using deep learning platforms to process large amounts of data.The traditional image processing technology,combined with deep learning to further solve practical problems is already the trend of artificial intelligence development,with high research value and application prospects.Optical Character Recognition(OCR)is a method of converting printed or handwritten characters taken by electronic devices such as cameras or scanners into text.With the development of character recognition technology,OCR has become one of the important applications in the field of pattern recognition and character recognition in documents has achieved a high recognition rate.However,for the character recognition of complex natural scenes,there are still some challenges,mainly because the character pictures taken from natural scenes have problems such as low resolution,out of focus,fuzzy background contrast,and too much interference.For example,in the application scenario of engineering drawings,complicated lighting problems,deformation of drawings,discoloration,obstructions,smudges and other interference factors bring great challenges to the positioning and identification of drawings.As Chinese national laws stipulate that architectural drawings can only be constructed and applied with the relevant state departments confirmed,only the drawings confirmed by the seal can be both convenient to search and can be applied to practical work.Based on the traditional image processing technology,this paper proposes an engineering drawing character sequence recognition system based on traditional image processing and deep learning.The system can take the stamped and confirmed engineering drawings through the mobile phone and take advantage of the improved RANSAC-table method to automatically classify different drawings.According to different classification results,it locates the specific geometric feature areas of the image and extracts the information in the areas.Subsequently,character features were extracted using a modified VGG(Visual Geometry Group)convolutional neural network.The bidirectional recurrent neural network BLSTM was used for feature decoding.The connectionist temporal classifier CTC was used to identify the end-to-end character sequence and intelligently extract drawing information.After being stored in the database,the rapid retrieval of engineering drawings can be achieved,reaching the goal of system intelligence and paperlessness.The innovations in this article are:1)Aiming at engineering drawing problems,traditional image processing methods are used to classify and identify drawings,and an end-to-end character recognition system for engineering drawings based on mobile devices is designed.Intelligent classification and identification of drawings can be realized,and it has advantages of intelligent,automated and paperless when implied on engineering application.2)In the engineering drawings classification of our system,combined with the Monte Carlo idea and random sampling consensus RANSAC,an improved RANSAC-table algorithm is proposed,which can effectively solve the problem with deformation,ambiguity and defect tables when addressing table detection.According to the comparison with traditional table detection algorithms,the proposed algorithm is more robust against engineering drawings.3)In the end-to-end character sequence recognition of our system,an algorithm combined with improved VGG convolutional neural network,bidirectional recurrent neural network BLSTM and connectionist time classifier CTC are used to solve the problem of character sequence recognition.Compared to traditional character recognition methods,there is a big increase in recognition performance.