Research On Text Detection Technique In Natural Scene Images

Text in natural scenes directly conveys high level semantics,thus is a kind of key element to scene understanding.With the rapid growth of images and videos,detecting text in natural scene images have gained increasing attention from the computer vision community.However,text detection in natural scene images is extremely challenging.On one hand,texts in natural scenes have different fonts,scales,orientations,colors and even different languages;on the other hand,various factors,such as noise,blur,highlight and partial occlusion,all may make it difficult to detect text in natural scenes.So,there are some technical difficulties for text detection in natural scene images.Considering variations of scene characters and the complexity of background,this thesis investigates the basic issues in scene text detection,by combining attest advances in image processing,object detection,pattern classification and machine learning.Concretely,the research contents in this thesis are as follows:1)Uneven illumination,characters distortion,partial occlusion in natural scene image and edge detection algorithm,all of these factors,have great impact on performance of text detection.They make contributions to absence of partial stroke in character detection,and even lead to false positives,missing detection.Aiming at these problems,this paper proposes a text detection algorithm based on stroke width by incorporating color cues of text pixels.In the process of character extraction based on stroke width,exploiting stroke color to extend stroke connected components,can resolve the trouble of missing stroke and false positives.On the foundation of character connected components,using the geometric properties and color consistency of adjacent characters can deal with missing characters in character detection well.The experimental comparison in each process of algorithm,proves that the text detection algorithm based on stroke and color fusion is effective to remedy the deficiency of the stroke width based detection algorithm.2)Blurring text and low-contrast text are easily confused with background,which makes difficult to obtain clearly distinguishable textual feature from natural scene image.Focused on these problems,a text detection algorithm based on extremal region is proposed.Firstly,extremal region operator is employed to detect character connected components in multichannel.Then characters can be generated after filtering non-character connected components and recalling character connected components.Lastly,a concept named text entropy is proposed by combining the number and the category of characters in line.Moreover,text entropy and convolutional neural networks would discriminate candidate text together.The experimental results show that extremal region based text detection algorithm can detect blurring text and low-contrast text effectively.3)Due to complexity of background in natural scene image,character detection has bad performance.What’s more,performance in text detection could be impacted.At the same time,the detection algorithm present has a limit to arbitrary oriented text.Consequently,a text detection algorithm based on convolution neural network is proposed to handle with above problems.To avoid character detection has great effects on text detection,a word is used as detection unit.The parameters of scale and ratio are reset in the region proposal network according to word.In order to obtain powerful text feature in text region,the local feature and global feature of different convolutional layers are integrated.Adding the angle of text in the multitask process of identifying candidate regions and location regression,makes model has the ability to predict geometric coordinates and angles of text area.Finally,the Monte Carlo based non-maximal suppression method is used to eliminate the redundant detection results.The experiment proves that the text detection algorithm based on convolution neural network realizes arbitrary oriented text detection in complex environment.