归一化互信息量最大化导向的自动阈值选择方法

当前景或背景的灰度分布呈现为非正态分布特征时，比如极值、瑞利、贝塔或均匀分布，将所选阈值与最优阈值之差控制在10个灰度级内并非易事.为了在统一框架内处理不同灰度分布情形下的阈值选择问题，提出了一种归一化互信息量最大化导向的自动阈值选择方法.该方法先采用多尺度梯度乘变换规范化输入图像，获得具有单峰长拖尾灰度分布的规范图像；然后对不同阈值对应的二值图像进行轮廓提取，获得不同的轮廓图像；最后计算规范图像和不同轮廓图像之间的归一化互信息量，并以最大值对应的阈值作为最终阈值.在具有不同灰度分布模式的9幅合成图像和59幅真实世界图像上，将提出的方法和1种人工阈值方法及4种自动阈值方法进行了比较.实验结果表明，提出的方法虽然在计算效率方面不优于4个自动方法，但在分割的适应性和精确度方面优势明显:对前述不同灰度分布情形，其所选阈值与最优阈值之差都在9个灰度级内.

Automatic Threshold Selection Guided by Maximizing Normalized Mutual Information

When the gray level distribution of foreground or background appears a non-normal distribution, such as extreme value, Rayleigh, beta or uniform distribution, it is challenging to keep the difference between the selected threshold and the optimal threshold under control within 10 gray levels. To deal with the issue of threshold selection in different gray level distributions within a unified framework, we propose an automatic method of threshold selection that is guided by maximizing normalized mutual information. Firstly, this method applies a multiscale gradient multiplication to an input image, which produces a normalized image with a single-peaked trailing distribution. Then, the binary images corresponding to different thresholds are subjected to contour extraction to produce different contour images. Finally, normalized mutual information between the normalized image and different contour images is calculated and the threshold corresponding to maximum is taken as the final threshold. The proposed method is compared with a manual thresholding method and 4 automatic ones on 9 synthetic images and 59 real-world ones with different distributions. The results show that the proposed method is not superior to these 4 automatic methods in computational efficiency, but it has a significant advantage in the adaptability and accuracy of segmentation:in the case of the aforementioned different gray level distributions, the differences between its selected thresholds and the optimal ones are all within 9 gray levels.