Research On Casting Identifier Detection And Recognition Under Low Contrast Based On OCR Technology

In recent years,optical character recognition(OCR)has been widely used,and its related algorithms are becoming more and more mature in character detection and recognition In the industrial scene,the identification character on the casting workpiece plays a decisive role,through which the detailed information of each link in the whole casting process can be queried.If OCR technology is used to automatically identify identity characters,it will greatly reduce people’s work and enhance work efficiency.However,in the case of low contrast between background and character,the traditional OCR technology has the many problems.Therefore,it has important practical significance to study the detection and recognition technology of casting identification characters under low contrast.The main work of this paper is to establish the casting identification character detection and recognition framework by deep learning,and design and implement the OCR system.OCR steps are mainly text area detection,character recognition.In the casting process,the identification characters are blurred after the white die is painted and the casting is painted,and the contrast between the background and characters is low,which leads to the increased difficulty of detection and recognition.In this paper,deep learning is combined to study this problem.The specific content is as follows:(1)Summarized the relevant knowledge of OCR technology.The neural network related technology in deep learning theory is expounded.(2)on the characters of the detection problem,based on the collected real data to correct and grayscale processing,using Text detection based on connected Proposal Network(Connectionist Text Proposal Network,CTPN)algorithm to test the casting identification character,through the experiment,compared with the traditional detection algorithm,the CTPN data sets the value of F-Score is0.73,It completes the task of casting identification character detection under low contrast well.Finally,the detected characters are clipped to facilitate the completion of subsequent recognition tasks.(3)on the issue of character recognition,the use of a Neural Network based on convolution cycle(Convolutional Recurrent Neural Network,CRNN)and connection time domain Classification(Connectionist Temporal Classification,CTC)end-to-end optical character recognition model,the model based in convolution layer VGGNet16 simplified and improved,and by using rotating,add noise,adjusting brightness and contrast data augmented method to solve the problem of the sample number less,then,the appropriate model parameters were selected according to the experimental comparison to realize the identification character recognition of casting workpiece under low contrast.Experimental results show that the improved character recognition model is stable,has high recognition rate,and has good robustness to similar characters.Finally,a complete OCR model is proposed by combining the casting identification character detection module and recognition module.In the experimental comparison,compared with other algorithms,has a certain improvement in identifying the identification characters of casting workpiece,and has a better performance.