Research And Implementation Of Scene Text Detection And Recognition Based On Channel Attention Mechanism

Scene text detection and recognition is one of the important research fields of computer vision,which has great significance in practical application.To recognize the scene text,we first need to locates the text instance through the scene text detection algorithm,and then recognize the detected text area.At present,the difficulties of scene text detection and recognition mainly include:(1)The inaccurate boundary of irregular scene text and text misrecognition caused by the complex scene;(2)The poor robustness of multi-scale scene text detection;(3)The attention drift caused by the low-quality image.These problems degrade the performance of scene text detection and recognition algorithms,and affect the practical application in real life.In order to address these problems,this paper focuses on irregular scene text accurate boundary detection,multi-scale scene text detection,and complex scene text recognition.The main work in this paper is as follows:1.Scene text detection algorithm based on feature enhancement pyramid network.Aiming at the irregular text inaccurate boundary and the lack of robustness to multi-scale text detection,this paper proposes a scene text detection algorithm based on feature enhanced pyramid network.The feature enhancement pyramid network adds a down-sampling path on the basis of the feature pyramid network,and only fuses the adjacent hierarchical features to avoid error accumulation.Then the channel attention mechanism is used for feature selection to enhance the features related to the text,so as to make the text boundary more accurate.Experimental results show that the proposed algorithm improves the detection F-measure by about 2%,which is better than most methods in recent years and achieves the state-of-the-art performance on MSRATD500 and CTW1500 datasets.2.Scene text recognition algorithm integrating channel attention mechanism.Aiming at the attention drift caused by the low-quality image and text misrecognition in a complex scene,this paper proposes a scene text recognition algorithm integrating channel attention mechanism.Firstly,we use the super-resolution module to improve the quality of image.Secondly,the channel attention mechanism is integrated into the encoder to strengthen the text features and reduce the background features.Finally,the spatial location information module is added to the decoder to enhance the global information,reduce the impact of attention drift,and improve the recognition accuracy of irregular text.Experiments show that the average recognition accuracy of the proposed method is improved by 3%,which outperforms most other algorithms in recent years and achieves competitive performance over the state-of-the-art works.3.Designed and implemented a scene text detection and recognition system.Based on the scene text detection and recognition algorithm proposed in this paper,we design the scene text detection and recognition system and implement it by using JavaScript,Java,Python,and other related technologies.All the data involved in the system is saved by the MySQL database.Finally,we complete the scene text detection module,scene text recognition module,and user management module.In this paper,the detection accuracy of scene text can reach 85%,the recognition accuracy of regular scene text can reach 94%,and the recognition accuracy of irregular scene text can reach 85%.To sum up,this paper proposes a scene text detection algorithm for irregular text,multi-scale text and a scene text recognition algorithm for complex scene text.Which makes the irregular text boundary detection more accurate,improves the robustness of multi-scale text detection and the recognition accuracy of scene text in a complex scene.