Research On Recognition And Extraction Of Text Information Under Complex Background

Text information in natural scene images contains rich and accurate high-level semantic information,which is the key to our understanding of the content elements of the scenes.With the development of technology and the improvement of living standards,the application of images and videos has grown rapidly.Therefore,text detection and recognition technology in natural scenes has received widespread attention from researchers.However,the complexity of natural scenes and the influence on uncertain factors is exactly a bottomnect,which brings great difficulties to this field.There are still many research issues for text detection and recognition in natural scenes.thus,this paper works on the above mentioned problems.It is based on the latest development in the fields of image processing,object detection,and machine learning.The detailed works can be briefed as follows:(1)Since images in the natural scenes are affected by uneven illumination,diverse character forms,partial occlusion,image noise,selection of edge detection algorithm and other factors,the performance of traditional SWT algorithm and MSER algorithm is greatly limited.They lead to missing detection in test results,or even false positives.Aiming to the shortcomings of the above algorithms,a text detection algorithm based on SWT and multi-channel lighting equalization MSER is proposed.First,the input image is denoised and corrected.Then light equalization is used to process the text image under R,G,and B channel respectively.The results are used to pre-screen the text area using the mser algorithm.Finally,based on the SWT algorithm,a SVM classifier is used to filter out the pseudo-character area to obtain the final text area.Experimental results show that the algorithm can effectively solve missed detection issues during text detection.(2)Classical Faster R-CNN algorithm is suitable for some conventional object detection tasks such as pedestrians.With regard to text detection,it neglects the particularity in the text that only focuses on the depth characteristics of the characters.Besieds,it ignores the context existing among characters in the text characters,thus results in poor text detection.Therefore,this thesis introduces a two-way LSTM network to extract the context information of characters based on Faster R-CNN.For predicting the geometric coordinates of the text and the tilt angle of the text,the angle information is added for judging and position regression in the candidate area.Moreover,Monte Carlo non-maximum suppression method is introduced to filter the redundant detection results.Finally,the method is experimentally verified to detect text in any direction in natural scenes effectively.(3)Different from conventional algorithms that divides detection and recognition into two parts and processed separately.This thesis proposes an end-to-end text recognition system based on CNN+BLSTM+CTC,which combines text detection and text recognition,and outputs the results together.The proposed algorithm adopts CNN and BLSTM as encoders,and connects CTC as decoders for text recognition.Compared with traditional detection and recognition methods,the proposed recognition method is more effective in recognition rate and the runtime is more efficient.