目的 在光学字符识别中,文本图像经常会出现一定角度的倾斜。为将倾斜的文本图像校正,以便于字符识别中的后续处理,快速准确地检测倾斜文本图像的倾角是非常重要的。方法 对基于投影轮廓的算法进行改进,提出了一种两级投影直方图方差的算法(TPHV)。首先在预定的角度范围内以一定角度步长对选定的图像区域做多方向投影,以获取投影直方图;然后计算各角度投影直方图的均方差,求出所有投影直方图方差的最大差分,将对应的投影角度作为倾角的粗略估值,最后以粗略估值为中心,以第1次投影步长为半径的角度范围内,再次以给定的检测精度为步长进行投影,重复第1次投影的工作,求出投影直方图方差的最大值,以对应的角度作为图像倾角的检测值。结果 该算法能够处理各种复杂的文本图像;对于诸如2 480×3 508像素的较大图像,可在200 ms左右的时间内完成倾角的检测;可检测的倾角范围不受限制;对相关网站提供的5组共500幅测试图像检测误差绝对值均值不超过0.5°,最大值不超过0.7°,检测误差的方差不超过0.1。结论 实验结果表明,该算法具有明显优势:速度快,倾斜角度检测精度高,误差集中,检测范围大,对噪声不敏感,具有广泛的适用性,适合于复杂的排版方式。

Objective In optical character recognition, skew is typically introduced into the document images obtained during the process. Fast and accurate skew detection is important to implement skew correction to these tilted document images, and thus, facilitates subsequent processing. Method An improved projection profile-based approach,called two-stage projection histogram variance, is proposed in this study. Angle space is discredited at a certain step length within the scope of a predetermined value.Projection histograms of the number of dark pixels are obtained at each possible angle.Variances of all histograms and their maximal difference values are calculated.The angle that corresponds to the maximal difference is selected as a rough estimation of the skew angle. New histograms are computed in the same manner, but angle space is discredited at the increment of the detection precision between the sum and difference of the rough skew estimation value and the step length used in the first histograms.The maximal value of the variance of the histograms is calculated.The corresponding angle is calculated as the final skew angle estimation. Result The proposed algorithm can be applied to all kinds of complex document images. The mean and maximal absolute values of error resulting from the algorithm do not exceed 0.5° and 0.7°, respectively. The maximal variance of error does not exceed 0.1.Thus, the proposed algorithm exhibits the most concentrated error distribution compared with other methods. Furthermore, the processing speed of the proposed algorithm is fast. Skew detection for a document image with 2 480×3 508 pixels can be accomplished within 200 ms by the algorithm. Conclusion Tests results show that the proposed algorithm exhibits fast running speed, high precision and wide scope for skew detection, strong resistance to noise, and excellent adaption to complex document layout.