Research And Implementation Of Text Detection And Recognition Sytem Based On Deep Learning

Text detection and recognition technology is a basic technology in the field of computer vision,which has been widely used in all kinds of image processing systems and has strong research value.However,there are still some problems in various types of text detection and recognition systems:1)the text detection methods for open scenes are usually unable to solve the problem of multi-form text modeling and overlapping instance detection at the same time.A few two-stage text detection methods can alleviate this problem,but their speed performance is poor;2)Most of the current text detection algorithms are usually modified from the general target detection algorithms,which do not make full use of the inherent features of text instances;3)Text detection and recognition systems for restricted scenes do not make full use of the priori information of specific scenes,and often introduce too many user constraints,which reduces the availability of the system.To solve the above problems,this paper makes targeted improvements based on the deep learning method.The specific research contents are as follows:1)In the task of scene text detection,this thesis proposes a single-stage text detection algorithm with the fusion of regression and segmentation.Based on the single-stage object detection framework,the algorithm additionally introduces a dynamic convolution parameter generation branch and an image segmentation sub-network to achieve single-stage multi-morphological text segmentation and positioning.On the basis of retaining the advantages of the two-stage scheme,the algorithm has a simpler network structure and faster inference speed.On the IC17MLT data set,the algorithm can increase the F value by 0.7%compared with the two-stage scheme,and the inference speed increases by 22.6%.2)In the task of scene text detection,this thesis proposes a text detection algorithm based on label assignment and quality prediction to solve the problem that the current scheme does not make full use of the prior features of the text instance.The algorithm alleviates the problem of mismatch between the original scheme and the text instance through a label assignment scheme based on the centerline area constraint;the result of the quality prediction branch is used to help the merge algorithm judge the positioning quality,thereby alleviating the problem of inaccurate results of the original scheme;by adding the reverse feature fusion branch,the ability of multi-scale modeling of the network is enhanced.On the IC17MLT data set,the algorithm improves the F value by 2.0%compared with the previous algorithm,while only bringing about 9%additional reasoning overhead.3)In the study of text detection and recognition systems in restricted scenarios,this thesis focuses on the problem of high user input restrictions in card recognition scenarios,and proposes and implements a card recognition OCR system with high user freedom.The system proposes a card card region positioning module based on segmentation to enhance the system's ability to process complex images;proposes a noise-sensitive text recognition module to improve the system's recognition ability in high-degree-of-freedom input scenarios.The proposed system can obtain reliable recognition results without restricting the user's input of background,posture,and angle.