Research On Text Detection And Recognition Algorithm Based On CNN In Natural Scene

Natural scene text detection and recognition technology has been increasingly popular in research area and has broad applications in computer vision,such as automobile scene character recognition,Cards and documents recognition,bill recognition,education scene character recognition.With the popularization of intelligent hardware devices,natural scene text detection and recognition technology has been widely concerned.However,due to the complexity of the scene text image background and the variability of the text,it is difficult to accurately detect and recognize the text.In this paper,the algorithm of text image detection and recognition in natural scene is studied as follows:(1)This paper introduces the research status of natural scene text detection and recognition algorithm,and analyzes the challenges in scene text detection and recognition.In addition,the common algorithms are introduced.(2)Aiming at the problem that CTPN network can only detect horizontal and slightly tilted text,improved multi-directional text detection algorithm based on CTPN is proposed and applied to multi-directional text detection and recognition system.By rotating the detection image at different angles,the CTPN network detects the initial position of the text in the rotated image.Then the candidate text boxes are fused to find the best text box.In the process of text box fusion,the text box fusion strategy is used to generate a rotating rectangular box to mark the text area.Experimental verifications are conducted on IC15 dataset,including detailed analysis for rationality of the algorithm based on CTPN.It solves the problem that CTPN network can only detect horizontal text and slightly inclined text.(3)A improved text recognition algorithm based on CRNN is proposed.The algorithm adds an adversarial network branch on the basis of thetraditional CRNN model.At the same time,the traditional text features and depth features are fused,so that the occlusion text can be recognized.The proposed CRNN based text recognition algorithm is trained on Mjsynth dataset,and tested on IC13 and SVT datasets.Compared with CRNN algorithm,the accuracy of the above two datasets is improved.Especially on SVT dataset,the accuracy of the algorithm without constraint dictionary is improved by 0.065.