Research On Text Detection In Natural Scene Images

With the rapid development of intelligent transportation, wearable device, the natural scene image analysis technique has been the current research focus. An important prerequisite for scene analysis of the image content is the text detection and localization technologies. Text’s character plays an important role on expressing visual information, and it’s also an effective clue to describe and understand scene content. Extracting text in the natural scene image will be widely used in the future on blind navigation, robot vision and other scenes. This paper mainly focuses on text detection and localization algorithm of natural scene. Based on Maximally Stable Extremal Regions(MSER) and stroke width transform algorithms, a MSER region detection method based on morphological filtering and an improved stroke width transform algorithm are proposed in this paper.Firstly, this paper describes the definition and properties of the MSER algorithm, and then contrast enhancement based MSER is introduced. By analyzing its sensitivity to blur the image, morphological filtering based MSER region detection method is proposed. In order to solve the problem of sticking character,morphological filtering method is used to improve the contrast enhancement based MSER algorithm. Gradient magnitude is used to enhance the grayscale image’s boundary in this method. Eight structural elements are designed in this paper. While performing morphological dilation on the edge pixels, the dilation structural element is chose according to the gradient direction of edge pixels adaptively. This method effectively solves the problem of character undetected and sticking due to image blur.Then, this paper analyzes the algorithm of stroke width transformation. In view of the strong dependence on the edge detection of stroke width transformation algorithms, edge-preserving filter is adopted on preprocessing stage. The constraints of edge points pair is modified as the gap appears easily on the cross strokes of the character. Meanwhile the connected component of character is generated with the color information of the image. The improved algorithm of stroke width transform is able to maintain the integrity of the character region and reduce the generation of non-text connected component while reducing background interference on character region.This paper designs a connected component based framework of text region location. Morphological filtering based MSER and the improved algorithm of stroke width transform are used to detect candidate character region in this framework respectively. Preliminary verification and SVM(Support Vector Machine) based verification is used to removed non-character connected components. Text line formation is performed by meanshift algorithm. Finally the layout and geometry of the text line is analyzed to segment the words. The framework of text region location is demonstrated on ICDAR2011 dataset, the results show that both the morphological filtering based MSER and the improved algorithm of stroke width transform perform well and achieve good results.