Research On Multi-features Natural Scene Text Detection Based On Image Enhancement

With the continuous maturity and popularization of intelligent terminals and the development of network technology,the mobile Internet is gradually integrating people's life.Text messages are being recorded and disseminated widely in the form of pictures.Therefore,extracting text from the image really provides a great convenience for people to receive environmental information.As the basis of extracting text information,the robustness and accuracy of text detection will directly affect the subsequent text recognition and image application.Compared to the traditional text documents which have a simple background,there are more interference factors in the natural scene picture,which brings a lot of uncertainty to the detection.Therefore,the subject of natural scene text detection has a certain challenge and practical value.The main research work in this paper is as follows:(1)Image enhancement and candidate region extraction.The text selection is using Maximally Stable Extremal Regions(MSER)algorithm for text candidate region selection.By analyzing the limitations of traditional MSER,it is found that MSER has poor detection performance for low-contrast images,and only detection of gray-scale channel will lose a lot of image information.Therefore,an image evaluation method is proposed in this paper,according to the gray distribution and edge projection determine whether the image belongs to low contrast map,and take the contrast enhancement processing to the low contrast graph.Using RGB and Perception-based tiny Illumination Invariant color channels and significant characteristic figure to replace gray level to achieve multichannel MSER detection,can make full use of the color information of image and eliminate the influence of small obstructions.(2)Extract and classify the feature of the candidate region.Extract Local Binary Pattern(LBP)features and Histogram of Oriented Gradient(HOG)features.Aiming at the limitation of texture features,this paper uses the relevant features of strokes--the number of boundary points and the area ratio of strokes.In view of the problem of inaccurate detection of the last two features in practical application,the corresponding improvement is carried out,also we improved the classification effect.We are using AdaBoost and SVM classifier to perform the performance test for each combined feature.We also do the performance and classifier test for the combined feature.Select the best feature combination and classifier.(3)Merge multichannel information and merge text rows.The multi-channel MSER causes candidate region to be distributed in each channel.There are differences in the candidate regions of each channel.And a multi-channel information fusion method is adopted for this problem.After multi channel merging,a single character corresponds to multiple candidate regions,which seriously affects text line aggregation.Therefore,candidate regions are reprocessed by overlapping types and color change rates.Aiming at the proximity principle and similarity of characters in text rows,the constraint conditions are set up.Combined with the idea of Hough Transform,the character quantity is merged into a text bank,and the final detection result is obtained.With using ICDAR2015 open database to detect the algorithm,the method gets a 79.3% accuracy and 72.8% recall rate,and the detection effect was improved.