改进的基于区域合并的纹理图像分割方法

Improved textured image segmentation method by region merging

针对自然纹理图像的特点,提出了一种改进的基于区域合并的纹理图像分割方法.首先选择符合人类视觉对颜色的感知区分能力的L*a*b*颜色特征;然后提取图像的Gabor能量作为纹理特征;接着由颜色相似度和纹理相似度的概率加权平均获得2个相邻区域的相似度;最后利用基于最大相似度的区域合并算法交互式地完成图像分割任务.实验结果表明:该方法比仅使用红绿蓝(RGB)颜色特征的相似度测量获得了更加精确的分割效果,并且在相同的初始过分割以及人工交互条件下,优于Lazy Snapping.

According to the characteristics of natural textured images, a novel image segmentation method based on region merging was presented. The color feature of L＊ a＂ b~ color space which accords with the color resolution of human visual perception was chosen. Then Gabor energy was extracted as a texture feature, and the similarity of two regions was measured by using the probability- weighted average of L＊ a＊ b＊ color similarity and texture similarity. Finally, the maximal similarity based the region merging algorithm was used to partition the image with the guide of user inputs. The experimental results demonstrate that the method can obtain better the segmentation accuracy than those only utilizing RGB color feature. Under the same conditions of initial over-segmentation and user interaction, the method outperforms the currently popular image cutout tool--Lazy Snapping.