Research On Scene Image Text Detection Based On Deep Learning

In natural scene images,text is the most common object,and it often appears on traffic signs,product packaging,and other objects.Effective detection of scene texts can help many applications implement specific functions.For example,an image-based geographic positioning system can implement the positioning function by detecting and recognizing the text of a scene image.Inspired by rapid development of deep learning,more and more object detection frameworks based on deep learning are used to perform text detection of scene images.However,since the text object in the scene image is different from the general object,the text object not only have the characteristics of rich scale,arbitrary orientation,and extreme aspect ratio,but also are easily disturbed by similar text backgrounds.Aiming at the distribution characteristics of text in natural scene images,this paper studies the text detection methods in natural scene images.The main work of this paper is as follows:(1)The scene image text detection model designed in this paper adds a text area detection module based on the SSD object detection framework.The module can adjust the default according to the difference of the feature map detected by the feature extraction layer of different scales.The aspect ratio of the preselected box,the shape of the convolution filter,and the spatial density of the default prediction frame.The module can efficiently combine the text detection results on each feature map,thereby enhancing the robustness of text detection at different scales.(2)The scene image text detection model designed in this paper is aimed at the SSD object detection algorithm which can only generate horizontal rectangular object area boundary detection frames,which leads to the problem of poor multi-directional text detection.Based on the SSD algorithm,a kind of Calculation strategy for bounding box of text area in any direction.So that the text detection model proposed in this paper can detect text in any direction.(3)In this paper,an end-to-end scene image recognition model is designed.Based on the scene image text detection model proposed in this paper,a text recognition branch and a bilinear interpolation sampling module composed of spatial transformation network are added.In addition to this,the idea of shared convolution features has been introduced so that the model can be trained end-to-end.The model not only can complete the text detection and text recognition tasks at the same time,but also can fully play the highly correlated and complementary relationship between the two and further improve the accuracy of the natural scene image text detection task.