Abstract: Aiming at the weakness of image detail and the huge calculation of the three-dimensional Otsu, an improved three-dimensional Otsu image segmentation algorithm is presented. A three-dimensional histogram based on gray level-average gray level-gradient is built. The original three-dimensional Otsu algorithm is decomposed into three one-dimensional Otsu algorithms. Based on the one-dimensional Otsu algorithm, a novel maximum scatter difference discriminant criterion is proposed. The criterion considers simultaneously both between-class distance and within-class distance. The experimental results show that the proposed algorithm does not only get satisfactory segmentation result, but also improves the calculation speed, which spends 900 times less than the recursive algorithm for the three-dimensional Otsu method.

Key words: image segmentation, Otsu, three-dimensional histogram, threshold

摘要： 针对三维Otsu图像分割算法运算量大和图像细节分割不够准确的问题，提出一种改进的三维Otsu图像分割算法。该算法基于灰度-邻域平均灰度-梯度的新的三维特征模型，构建三维直方图；将三维Otsu分解为三个一维Otsu；在同时考虑类间距离和类内距离的基础上，提出一种新的最大散度差准则。实验结果表明，该算法不仅可以取得较好的分割效果，而且计算量较小，比三维Otsu阈值分割递推算法快900倍左右。

关键词: 图像分割, Otsu, 三维直方图, 阈值