Research On Text Detection And Recognition Technology In Natural Scene Image

The text in the natural scene image contains rich and accurate high-level semantic information,which has guiding significance for blind obstacle navigation system,intelligent urban traffic management system,automobile driverless system and instant translation system.Therefore,it is of great research value to locate and identify the text in the natural scene image.This paper has conducted in-depth research on text localization and recognition in natural scenes.The specific work is as follows:This paper has conducted in-depth research on text localization and recognition in natural scenes.The specific work is as follows:1.The traditional MSER algorithm is sensitive to illumination and text missed detection occurs when text candidate regions are extracted on a single grayscale channel.To solve this problem,this paper proposes an MSER algorithm based on multi-channel illumination equalization.Firstly,the images are illuminated and equalized under R,G and B channels respectively;Then,the MSER region of the text character is extracted by the MSER detection operator under the corresponding channel;Finally,the MSER region of each channel is merged as the character candidate region.After experimental verification,the improved algorithm can detect relatively complete character regions for images with uneven illumination or different complex backgrounds,which improves the recall rate of the algorithm.2.The traditional MSER algorithm will cause false detection in scene text detection with complex background.To solve this problem,this paper proposes a pseudo-character region filtering algorithm based on multi-feature fusion.Firstly,the HOG feature,the LBP feature,and the CNN feature are respectively extracted for the character candidate region;and then the three features are serially fused;Finally,a character discriminator is trained by the SVM to filter the pseudo-character region.It is verified by experiments that this algorithm can eliminate more pseudo-character regions and improve the accuracy of the algorithm.3.The sliding convolution character model is based on the recognition of character classification.It only focuses on the depth characteristics of characters,but ignores the context relationship between characters in the text line,which greatly reduces the recognition accuracy of the algorithm.To solve this problem,this paper deeply studies the sliding convolution character model,and introduces the bidirectional LSTM network to carry out scene text recognition.First,remove the classification of the sliding convolution character model layer;then use the CNN sliding window to extract the characteristics of the input image sequences,the output sequence characteristics of the input to the design good two-way LSTM network to extract context features of each character;and finally,using the CTC transcriptional mechanisms transcribes LSTM output prediction for the actual string.Compared with other algorithms,the proposed algorithm can improve the recognition accuracy significantly.