Multi-Directional Chinese Text Recognition In Natural Scenes Based On Deep Learning

The text information in the image can provide important help for people and computers to interpret the image scene.Text recognition of images in natural scenes is one of the most popular research fields.Text recognition in a natural scene generally involves two steps of detecting a text area in the image and text recognition of the text area.This paper mainly studies the text detection and text recognition algorithms in natural scenes,and finally connects the detection and recognition to realize the end-toend system of text recognition.This paper proposes a multi-directional text detection method based on improved Faster RCNN for text detection in natural scenes.The SE module is added to the feature extraction network VGG16 of the detection algorithm to improve network performance.For the regression of the bounding box,the angle information of the text block rectangle is added,and the multi-directional text can be detected.For the multi-directional feature of the scene text,an angle variable is added to the anchor selection to generate more anchors.Finally,the detection algorithm proposed in this paper is tested on the open standard dataset.By comparing some existing detection algorithms,the proposed algorithm has better detection performance.This paper also studies the recognition method of Chinese text,and proposes an improved STN-CRNN method to identify the detected text area.STN-CRNN first corrects the text of the text box image and then sends it to the text recognition network for identification.The text recognition network performs feature extraction on the corrected text image,and generates a feature vector sequence.A attention mechanism is then introduced in the decoder module to decode the sequence of feature vectors to obtain the result of text recognition.