Character Recognition And Localization Of Engineering Drawing Component Based On High Order Convolutional Neural Network

With the rapid development of computer science and technology,its application in the field of industrial engineering is more and more extensive.In the actual engineering design,engineering drawings as an important part of the staff to carry out the construction of the information basis,and the key information into the drawings of the computer is also a more common work content.However,due to the large number of engineering drawings,the content is complex,the traditional artificial identification and input method of work efficiency is low,and the error rate is higher.Therefore,the art character recognition technology as a based on image processing,pattern recognition and artificial intelligence and other multi-disciplinary integrated application,in the field of engineering is of great significance.It has been around for 90 years since Optical Character Recognition(OCR)was put forwa rd.Optical character recognition usually includes two major parts: character localization and c haracter recognition.Although the character orientation has been mature,the efficiency and s pace requirement of the algorithm are gradually improved with the increase of image resolutio n and complexity.The traditional algorithm cannot get the accurate positioning in the drawing paper,and many algorithms have high accuracy in the resolution of up to ten 100 million or more times when the efficiency is low.Compared with the traditional feature extraction algori thm,convolution neural network can automatically extract the characters of the character,wit hout the need to re-transform the character and add the second one.In this paper,we propose a novel method for character recognition,which can be divided into two types: structural meth od and statistical method,which can reduce a lot of subjective selection errors,and show stron g robustness of the feature extracted by convolution neural network.Therefore,convolution n eural network has been widely used in the vision direction of image processing machine.So it is feasible to add convolution neural network.In this paper,we first select the relatively stable and high-accuracy connectivity domain a lgorithm based on the irregularity and unsaturation of English alphabets and digital structures,and compare the traditional localization algorithm with several mature localization algorithm s.Compared with the prior domain-based growth and contour-based connected domain algorit hm,the improved algorithm has improved the efficiency of region growing threshold connectivity domain.Considering the drawing features of the engineering drawings,we add algorithm s such as straight line detection,threshold judgment and size ratio determination to enhance th e experimental results.The research of convolution neural network based on convolution neural network is more abroad,and the research in China is not yet mature.With the increase of the complexity and t he variety of convolution neural network,the paper chooses appropriate image feature for diff erent target image Convolution neural network,improved neural network structure,selection of the optimal excitation function to enhance learning efficiency have also become important r esearch problems.Le Net-5 is an earlier convolutional neural network with relatively mature st ability in the initial recognition of handwritten digits.However,the recognition of letters and numbers in this paper is much larger than that of the original convolutional neural network.In this paper,the structure of the original convolution neural network layer,the output and the in centive function,pooling algorithm,and so have targeted improvements,saving the learning ti me and improve the accuracy of recognition.In order to ensure the validity of the experimental graph,30000 standard work drawings a re collected in the range of 3000 × 5000 to 10000 × 20000 in order to increase the complexity of the experiment.,This paper in the experiment by adding 1000 hand-painted engineering dr awings.The experimental results show that the accuracy of the proposed method is 97.8%,wh ich shows that the proposed improved connectivity domain algorithm is effective.The recogni tion rate is 99.4% and the Le Net-5 algorithm is 20% faster than the previous one.