Text Image Analysis Based On Deep Learning

This paper studies how to analyze handwritten text images and scene text images based on deep-learning methods.Text is the body of scanned document and handwritten text images.The background of these kind of iamges is clean.The region of text is easy to estimate,therefore,high level semantic researches,such as writing style,can perform.This paper analysis how network structure affects the accuracy of text independent writer identification.This paper learns how to design network structure according to the feature of problem to be solved.Finally,this paper propose DeepWriter network structure,outperform previous methodes a large margin.The identification accuracy is 99.01% when identifying 301 writers and 97.03%when identifying 657 writers according to one handwritten English text.The identification accuracy is 93.85% when identifying 300 writers according to one handwritten Chinese character.The research proves that analysing the style of handwritten text based on deeplearning methods is workable,laying the foundation for rearsh on high accurate writer identification on open set.The background of scene images is varible and complex.Text in scene images is not focused,but incidental.The text in scene images may be blured,small,uneven irradiated,distorted.Detecting text region accurately and robustly is the foundation of scene text recogniton.When text in scene is detected and recognized accurately and roubstly,research on high level artificial inteligence,such as visual navigation and scene understanding can be workable.This paper how to model text detection in complex background.This paper divide scene text detection into foreground segmentation,tob and bottom estimation,and,left and right estimation.According to the characteristics of these problem,this paper use different methods to estimate them.This paper conclude false positives and false negatives in scene text detection into sevel classes,and design two sets of rules for classifying error detections into these calsses automatically.According to the analysis result of error detections,the benchmark is suggested to be refined,in order to exactly evaluate the performance of algorithms,guiding the correct direction of research.The analysis result also suggest that the key point in scene text detection is how to estimate the start point and end point of text accurately.