Research On Metal Parts Identifier Detection And Recognition

The digitization,networking and intelligence of the manufacturing industry have become an irresistible trend.Computer vision technology is one of the indispensable technologies in the realization of industrial production automation.The production and inspection of machine parts are the basis of automated production,and part identifiers are an important factor in industrial intelligence for production intelligence,inspection intelligence,and control intelligence.A good text detection and recognition model can automatically extract industrial parts identifiers,thereby greatly promoting the intelligent process of industrial production.In industrial scenarios,texts have a more disturbing background,unique font style with concave,convex and other phenomena.And such texts account for a very low proportion of public datasets,which makes the existing models unable to perform well in the detection and recognition of industrial scene text.Based on deep learning,this paper detects and recognizes metal part identifiers.This paper mainly does the following work.(1)The related technologies and theories of text detection,text recognition and text spotting are briefly summarized,and the related technologies of deep learning are expounded.(2)The MSER+NMS algorithm was used to detect the identifiers of metal parts,and it was found that the traditional algorithm has shortcomings such as poor anti-interference ability,poor detection effect,and poor generalization ability.(3)For industrial scene text datasets that are rarely disclosed and the metal part identifier datasets are small,operations such as rotation,noise addition,adjustment of light and contrast,blur,corrosion,deformation,distortion,reorganization,and translation are performed.Image augmentation.(4)According to the requirements of precision in industrial production environment,a two-stage metal part identifier detection and recognition method with precision priority is proposed.In the two-stage method,the detection task of metal part identifiers is completed using the SAST model,and the detected text parts are cropped for subsequent text recognition tasks.In the text recognition problem,the CRNN model is used to complete the recognition of metal parts identifiers,and the MoblieNetV3 model is used for feature extraction.The experimental results show that compared with the traditional MSER+NMS algorithm,the two-stage method has fewer false detections in the detection task.When the MobileNetV3 model is used as the feature extraction network,the model converges faster and the model size is smaller,which is more suitable for industrial use.Scenes.(5)According to the requirements of speed in industrial production environment,a speed-prioritized one-stage metal part identifier detection and recognition is proposed.In the one-stage method,the detection and recognition of metal part identifiers are simultaneously accomplished using the PGNet model.The experimental results show that,compared with the two-stage method,although the one-stage method has lower detection and recognition accuracy,the speed is much faster than the two-stage method.And the one-stage method is used to verify the efficiency of the Fine-tune method.