Scene Text Detection Applied To The Blinds Visual Assistive System

Abstract:Text in natural scene images provides us with lots of important clues which are useful for many image-based applications, and the blinds visual assistive system is one of the most important applications. The system obtains surrounding environment by image acquisition devices; then extract image text information by text location technology; next recognize text strings by Opticm Character Recognition (OCR); finally output recognition results in the form of speech signals. The system transforms surrounding information from text to signal, so that it can provides the blinds with visual aids. A lot of computer technologies are used in this system, among which text detection is especially important. This is because exact text detection results have big influence on text recognition results and accurate description of the surroundings, which are meaningful for the blinds. However, natural scene images have complex background and the text from them is usually various, which make it an important and challenging work to detect text from natural scene images. In this paper, a text detection method based on color clustering is proposed, and mainly composes of the following four parts:(1) Color clustering. Text in same natural scene image usually has similar color, which can be used for text localization. In this paper, three color components are extracted to form feature vector and K-means is utilized to group pixels in images into five different color layers based on the feature vector.(2) Character identification. Color clustering results always contain text and a lot of scattered non-text components. In order to reduce the computation of character grouping and remove background noise, character geometrical and structural analyses are adopted to identify candidate characters.(3) Character grouping. In order to connect the scattered characters to form text blocks, the proposed method used the geometrical properties and space locations for character grouping.(4) Text verification. Candidate text blocks from character grouping are usually composed of true ground text and mis-detection non-text. In order to improve precision rate, text verification are used. In this paper, text verification is performed by geometrical and edge density identifications.The experimental results on the publicly available ICDAR2003dataset show that the proposed method can detect text in different natural scene images. Quantitative comparisons of the proposed method with other existing text detection methods have show the advantages of the proposed method.