Study On Stroke Width Transform With Geometric Constraints And Application In News Caption Text Location

With the development of mobile intelligent device, the demand of text information extraction in natural scene images increases sharply. The text information in natural scene images can be greatly helpful to social robots navigation and interaction, image retrieval and other fields. However, traditional OCR engine can only handle scanned documents and the results are bad when applied to natural scene images. The key issue lays on the text location, which is very challenging because of the mixture of text and non-text regions and the randomness of location of text. This paper studies the text location techniques in natural scene images. We improve the traditional stroke width transform with geometric constraint and global feature measurement of text area. On the basis of text location, we build a news captain location system.The main contributions of our work are as following:Firstly, based on the properties that text regions regularly present high visual saliency, abundant edges and consistent color. We detect the salience, vertical edge density to calculate the feature of candidate components and text lines in the whole image. Besides, we also convert the position and scale information of candidate text area to global measurement to compensate original SWT sensitiveness to local noise.Secondly, we utiliz geometrical constraints which combines the feature of the width, color and direction characterization of stroke to improve the tranditional SWT. In the process of SWT, we limit the formation of stroke, avoiding incorrect links between asymmetrical edge points. Compared with the traditional stroke width transform, our method performs better when edge is partly missing and is more robust to noise, blur and low contrast in natural images. By the means of geometrical constraints, our method clusters the more candidate text pixels in text regions and reduces the production of the non text components, which reduces the complexity at the character level and text line level filtering and avoids misjudging the ambiguous candidates.Based on the above research results, we apply the geometric constraint stroke width transform in the news caption location system in conjunction with the video image properties and the distribution rules of news caption text.