融合颜色和深度信息的图像物体分割算法*

当图像中存在阴影、低对比度边缘和模糊区域时，传统算法仅利用外观信息难以准确提取物体轮廓，而深度不连续性为辨识物体边界提供有用信息。文中提出基于颜色和深度信息的图像物体分割算法，首先利用mean-shift算法对图像进行适度的过分割，然后融合颜色和深度信息充分描述过分割区域的特性，根据深度信息自动选取目标和背景的种子区域，最后基于最大相似度进行区域合并，得到图像物体分割结果。在Middlebury和NYU-V2数据库上的实验表明，相比当前通用算法，文中算法简单有效，能提高分割的准确性，改善分割图像的视觉效果。

Image Object Segmentation Algorithm Combining Color and Depth Information

The object contour is difficult to be extracted by the existing methods only using appearance information in the image with shadow, low-contrast edges and ambiguous areas. The depth discontinuities provide useful information for object boundaries identification. An image object segmentation algorithm is proposed by combining color and depth information. Firstly, the image is over-segmented into small homogeneous regions by mean-shift algorithm, and then color and depth information are combined to describe the characteristics of regions adequately. Next, seed regions of target and background are automatically selected according to depth information. Finally, an object contour is extracted by maximal similarity based region merging (MSRM). Experiment results on Middlebury and NYU-V2 databases show that the proposed algorithm is simple and effective compared with state-of-the-art algorithms. Besides, it improves the segmentation accuracy and enhances the visual effect of the segmentation image.