Research And Implementation Of Scene Image Text Localization Method

Scene image often contain rich text information, which can provide important clues for many applications based on image, such as scene perception, auxiliary navigation and vehicle tracking, etc. So it has important value to complete text location and recognition in scene images.Scene image text processing system includes two important modules: text location and text recognition. As an important prerequisite for text recognition, the research and exploration of text location is of great significance. Although the research on the text area location technology has made some progress, it remains a huge challenge for scene image text location, duing to the complexity of background and all kinds of interference of external factors. This paper designs a new text location method and its main contents are as follows:1. The preliminary text location method based on the combination of the maximally stable extremal regions and the corner feature.Maximally stable extremal regions(MSER) feature has good robustness to the light and viewing angle changes, so this paper proposes a preliminary text location method based on this feature. Firstly, the paper analyzes problems when MSER detection algorithm is directly used in the scene images, it proposes to make use of histogram equalization technique and gray morphological operation to improve image, and extracts MSER feature in the pretreatment scene image. And then, connected component analysis is conducted for the extracted MSER feature, and candidate areas are filtrated by using heuristic rules. Finally, this paper presents the idea of filtrating candidate areas by combining the corner feature for the characteristics which text areas has rich corner feature, and finally the paper get the preliminary location result.2. The quadratic discriminant of preliminary location results.This paper puts forward to utilize the stroke width feature to make quadratic discriminant of preliminary location results as the problem that some non-text areas still exists. The paper analyzes the issue of incomplete feature of SWT, and proposes to solve the problem by combining image color information and text edge pixel gradient information. The paper further removes non-text areas by using stroke width features.3. Correction of text line location results.The paper gets text line location results by using text line construction algorithm for quadratic discriminant result areas. When the paper considers text areas that occur rotation, affine and projective transformation, the text line location results have some problems which causes the inaccuracy of location results: interference between overlapping text boxes, too many background noises contained in text areas and text deformation. The paper proposes to correct the text line areas by using low-rank decomposition technique, and finally it obtains accurate text location results.Finally, the paper gives part of the location and correction experiments, and compares the data with other algorithms. Experimental results show that the algorithm of this paper has a better adaptability for cxomplex background and environmental factors, at the same time, the paper completes the tilted text line location, and the location speed also has relatively improved.