关键词: 文本检测, 边缘增强, 最大稳定极值区域, 颜色空间, 贝叶斯模型

Abstract: In order to solve the problem of text regions can not be extracted well in low contrast images by MSER, a novel method based on edge enhancement was proposed. According to the prior knowledge, it can be known the gradient at the edge of the image is the largest, the contrast of images can be improved by edge enhancement, the correlation operation of original image and orientation gradient which obtained by HOG feature descriptor can effectively enhance the edge information of image. The color information can be lost when the original image is converted to grayscale image, however, the color space contains a wealth of valuable information. To make the improved MSER work better, the system applied MSER in HSI color space which satisfies human visual perception better, this improvement can extract more interesting text regions. Then, the Bayesian model was used for the classification of characters, the main three features of stroke width (SW), edge gradient direction and inflexion points were used to filter non characters. Finally, grouped the characters into text lines. The system demonstrates the performance on standard benchmarks, such as ICDAR 2003 and ICDAR 2013, the experimental results demonstrate that the method based on edge enhancement can extract more interesting text regions, it is robust to complex and low contrast images and can achieve a higher recall rate.

Key words: text detection, edge enhancement, Maximally Stable Extremal Regions (MSER), color space, Bayesian model