Research On Scene Text Detection And Recognition Based On Deep Learning

The text in the natural scene image has more accurate semantic information,accurately extracting and recognizing the text from the scene image can greatly enhance the computer’s ability to understand the image content.Computer detection and recognition of characters in natural scene images has become the most concerned research hotspot in the field of character recognition and even the whole field of computer vision.Therefore,based on the deep learning method,this paper makes the following research on text detection and recognition in scene images.Firstly,aiming at the problems of complex background,various visual elements and accurate detection of characters to be recognized in the scene image,a object detection algorithm based on deep learning and YOLOv5 is proposed to detect and locate the characters in the image,and the text detection experiments are carried out on various types of datasets.The results show that the object detection model can locate different types of characters in the image quickly and accurately,which provides a basis for the subsequent recognition work.Secondly,for the problem that the image background is easy to be confused with the text and affect the recognition,the semantic segmentation model based on UNet and the Att-UNet model embedded with attention mechanism are constructed,the text in the scene image is detected and located by using the method of image segmentation.The dataset of pixel level annotation is constructed by adding artificial noise and interference to the image,which alleviates the problem of lack of dataset.The experimental results of semantic segmentation show that the segmentation method can not only segment the text line area,but also directly segment the shape of the text itself,which greatly enhances the features required for subsequent recognition.Finally,for the problems of different text arrangement and different detection results in scene images,the recognition of single character,fixed length text line and indefinite length text line images based on deep neural network is studied respectively:Based on the modified classification model,the high recognition accuracy of single character images is realized,the amount of parameters is reduced,and the advantage of understanding the coupling model is proved;Based on multi-label classification method,an end-to-end fixed length text line recognition model is proposed;Based on CRNN model and CTC loss,the end-to-end recognition of indefinite length text lines is realized.The end-to-end text line recognition method avoids the error accumulation caused by the traditional text segmentation method,and is conducive to the improvement of the overall accuracy of scene text detection and recognition.