Natural Scene Text Detection Algorithm Research Based On Mobile Terminal

Natural scene contains a large amount of text information, so it is of great important to extract the text information from the natural scene images. Text information in the scene images has variable forms, and it varies in the sizes and directions. What's more, the background of the text in the natural scene images is complex. These factors bring great difficulties to the detection of text. The correct detection of the text in the scene has a direct impact on the accuracy of the subsequent text identification, so the correct detection of the text plays a very important role in the text processing system. The algorithm of this paper is completed on the integrated scene word processing system of our laboratory, which is based on the Android platform. Considering the limitation of mobile resources and the processing speed, this paper mainly studies two kinds of text detection algorithms for mobile phone. Experiments show the effectiveness of the algorithm, and this paper did some corresponding improvement from the perspective of the entire system.The specific work of this paper is as follows:First of all, this paper introduces the working process of mobile terminal natural scene text processing system, gives the results of each step of the operation in the system, introduces the concrete realization way of the navigation and positioning system achieved by this paper, and gives the results of the system.Then, this paper introduces the text detection algorithm implemented on the mobile terminal natural scene text processing system. In this paper, an improved text detection algorithm based on stroke width transform is proposed. The stroke width transform is fast and can be used for text detection on mobile phones. But the gradient direction of the text can not be determined in advance when using, so the correct stroke width chart can not be finished. In this paper, we improve the algorithm from two aspects, one is to get the correct stroke width by determining the direction of the text gradient, and the other is to get the correct stroke width chart by the way of the drawing width of the positive and negative direction. In this paper, the accuracy and recall rate of the algorithm is obtained with the performance evaluation system. The results show that these two improvements have achieved certain results.Finally, a text detection method based on the connected domain and the feature of corner points is presented. Although the first text detection method has achieved a certain effect, but application in practical is still not ideal. Therefore, this paper proposes a method for text detection based on the characteristics of connected domain and corner point density. First, the two value image is obtained by using color edge detection and morphological processing, and candidate text regions are obtained by using the connected domain analysis algorithm. Then, the corner feature of the image is extracted, and the corner density of each candidate text region is calculated. Finally, the text region is obtained by filtering out the connectivity region according to the feature of the corner density of the candidate text region. This method has low computational complexity, is simple and fast to process. After the experiment, the accuracy of this kind of scene text detection algorithm is improved by 6.7% and the recall rate is increased by 15.5%.