Research On Complex Scene Text Extraction Methods Based On Visual Saliency And Color

With the development of information technology, digital camera and smart phone with camera has become popular with people nowadays. The popularity of digital devices not only changes the life style and the culture of people, but also products a large number of digital images at the same time, in which there are a huge amount of information. Not only objects and scenes are contained by the information, but also text information does. The text information in the digital image is especially vital for people, but how to extract them is a challenge mission and a urgent problem to solve.The complexity of scene images is a common and difficult to solve interference factor for text extraction algorithm in scene images. This dissertation places importance on text detection and localization in the complex scene digital image, and designs two kinds of visual saliency based scene text background suppression algorithms(Visual Saliency and Boosting Based Background Suppression Algorithm for Scene Text, and Visual Saliency and Text Confidence Map Based Background Suppression Algorithm for Scene Text), and a sort of scene text localization algorithm based on HSL color space.The contributions and contents of the dissertation are listed as below:1. The scheme of background suppression for scene text is proposed, which combines with histogram of oriented gradient features, its statistical features, gradient magnitude features and gradient curve features based on the boosting method and the characteristic of visual saliency. Firstly, the scheme obtains the saliency map according to the spectral residual theory; and then averages the saliency map and the confidence map which is obtained from the cascade boosting calibrated classifier with the geometry mean method; finally, succeeds to suppress the background. The scheme aims to suppressing the complex background and highlighting the foreground text in natural scene, which could be as the preprocessing stage of the scene text localization algorithm, and improve the performances of the scene text localization algorithm. The experimental results on both the ICDAR 2011 text localization competition test dataset and laboratory scene Chinese text dataset show that the scheme can effectively suppress the complex background and improve the scene text localization algorithm.2. The method of background suppression for scene text is presented based on visual saliency and text confidence map. Firstly, the scheme obtains the saliency region according to the spectral residual theory; secondly achieves text confidence map in the saliency region with the fraction map of corresponding pairs based on stroke feature and the posterior probability map based on Fourier spectrum statistical feature; finally, saliency map, text confidence map and HSL color feature are combined with the graph model to suppress the background and highlight the foreground text. The experimental results on both the ICDAR 2011 text localization competition test dataset and laboratory scene Chinese text dataset show that the scheme as the preprocess stage can effectively suppress the complex background and improve the scene text localization algorithm.3. The corner-type feature based on photometric invariants and the Histogram of Oriented Gradients of Edge Magnitude(HOG-EM for short) statistical feature are proposed owing to the drawback of the text localization algorithm in non-text texture-rich regions. A two-stage multi-layer text localization algorithm in complex scene is presented On the basis of the two novel features. Firstly, edge map is obtained and 8 layers of binary maps in the HSL space are generated according to the characteristic of the HSL space in the proposed method. And then 9 layers of sub-maps are formed to gain text candidate components with multi-layer connected component analysis; secondly, the two novel features mentioned above are extracted to remove the non-text block from text candidate block and verify text. Experiments indicate that the proposed scheme can efficiently remove non-text texture-rich regions, decrease false-alarm rate and obtain reasonable accuracy and recall rate.