The Application Of Hue-level Segmentation In Text Detection Of Natural Scene

With the rapid popularization of universal digital camera,the rapid development of computer network and mobile Internet,the widely application of the intelligent mobile terminal and the increasing innovation of wearable devices, to record their daily work, study and life through images and videos has become a very active means. As the important semantic information in texts of natural scene images, the precise detection and recognition of these texts are particularly important for image understanding. Therefore, in academia and industry, it has attracted much attention of people.The difficulties of natural scene text detection are mainly caused by the diversity of the text, the uncertainty in the process of image acquisition and the scene complexity. Aiming at the problem that complicated backgrounds in natural scene have serious interferences on accurate text detection, a new method based on hue-level segmentation is put forward. A method based on color distribution is used to test its performance. And it is applied to SWT and MSER text detection methods to prove its performance. All tests achieve the expected effect.In the hue-level segmentation method, the images of RGB color space are transferred to YUV color space, and the hue-level is obtained by the orthogonal U and V channel signal. Using dual threshold OTSU method, the hue-level statistics histogram getting from the YUV color space is used to calculate the two segmentation thresholds. One image is segmented into several layers according to the two thresholds. The backgrounds complexity of each layer image is reduced and the detection effect will be improved. Using the space distance between two pixels, local color intensity distribution image will be obtained. Its density image is obtained by average filtering and its binary image is obtained by setting a simple binarization threshold of color intensity distribution image. The crude extract text is realized by these three images, and then text detection is realized by text heuristic rules.The segmented multi-layer images are applied in methods based on color intensity distribution, stroke width transformation and maximum stable extremal regions respectively. Experiments prove the validity and the better flexibility of this method and improvement of the recall rate and the detected text connected components.