Irregular Text Recognition Model Integrating Channel Attention Mechanism
Aiming at the problem of difficult text recognition caused by slanted text fonts and complex backgrounds in natural scenes,this paper proposes an irregular text detection and recognition model that incorporates a channel attention mechanism.First,the model combines the CRAFT detection framework and image post-processing technology to detect irregular text areas with the probability of character centers and adjacent character centers,and implements image background replacement,tilted text rotation correction and text line extraction in the detection area;then a text recognition model(SE-CRNN)incorporating with a fused channel attention mechanism named as SEnet is developed,which achieve text recognition in complex backgrounds by adaptively learning channel weights and employing data augmentation techniques.Experimental results show that the model still has high text recognition accuracy even when the text is tilted and grayscale is inconsistent.The recognition accuracy of the SE-CRNN model on the test set is as high as 84.2％.
text detectionimage post-processingattention mechanismdata enhancement
万方数据
