Research On Text Detection And Recognition Of Metal Plate And Strip Product Label Based On Deep Learning

The detection and recognition of product label text information on metal strips shall be generated in response to the needs of the steel industry for transportation operations,the use of intelligent methods instead of manual methods not only can improve the accuracy of information records,but also promote the construction of intelligent logistics,which has important research significance.The paper,based on the detection and recognition of metal strip product label text information,thoroughly studies the techniques of classical line detection,perspective correction,semantic segmentation based on deep learning,text detection and recognition,and proposes an end-to-end metal strip product label text identification solution that integrates correction processing,text detection,recognition and error correction,and completes the development of the demo system.The paper improves the defects of some algorithms,and conducts an experimental analysis to demonstrate the feasibility and effectiveness of the algorithm based on the open source dataset and the metal strip product label dataset.The core content of this paper can be summarized as follows:(1)For the perspective of the label area in the product label image of the metal strip taken on the industrial site,such as perspective,affine transformation and distortion,the paper implements the method based on linear detection and positioning correction,and used the linear discriminant analysis gray scale method to provide support for line detection,which improved the effect of line detection.In the case of poor generalization of traditional image processing methods,a method based on deep learning semantic segmentation and localization correction is proposed in a pioneering way,which lays a solid foundation for accurate text detection and recognition.(2)The paper delves into the theoretical approach of text detection based on deep learning.Through the theoretical research and experimental analysis of the EAST and Pixel Link detection algorithms,the advantages and disadvantages of the regression and segmentation ideas in scene text detection are discussed in detail.According to the characteristics of industrial scene image,Pixel Link algorithm is improved,and the experiment proves the efficiency and accuracy of the improved algorithm.(3)The paper delves into the CRNN based on connectionist temporal classification.In view of the characteristics of blur,occlusion,illumination and distortion of the text images obtained in industrial scenes,the original CRNN network is deepened and widened by the attention mechanism,and then CRNN algorithm is improved,which expands the ability of feature extraction and improves the ability of the algorithm to identify interference information.The obtained model achieves good recognition results.(4)Aiming at the particularity of the recognition scene,the method of regular split matching,similarity dictionary matching and hybrid matching are proposed to realize the retrieval and correction of key fields.In the end,the solution proposed in this paper is systematically developed based on the QT framework,an end-to-end method of the detection and recognition for the metal strip product label text is implemented.