少量交互的多视角图像目标分割算法

发布时间：2018-05-12 10:06

本文选题：多视角图像 + 目标分割　；参考：《计算机辅助设计与图形学学报》2017年01期

【摘要】：目前,单一图像在人工交互的条件下已能实现准确分割,但如果图像数量大,则需要大量劳动量.而自动分割难以在多图像间构建统一的目标模型.为此,提出一种多视角图像的少量交互分割算法,即仅对其中一幅进行交互,对其他图像进行前景目标一致的自动分割.首先通过构建特征点轨迹图模型,建立其他图像特征点与交互图像特征点之间的连接;然后利用最短路径算法进行特征点的前景/背景分类,获得多图像间的局部对应;最后将已分类特征点作为形状先验,使用测地星凸分割算法实现多视角图像的一致分割.命中率和误报率的比较实验结果表明,文中的算法具有较好的一致分割能力,而且这种局部对应的方式还可以实现多目标的分割.
[Abstract]:At present, a single image can be segmented accurately under the condition of manual interaction, but if the number of images is large, it needs a lot of labor. But automatic segmentation is difficult to build a unified target model between multiple images. For this reason, a small amount of interactive segmentation algorithm for multi-view images is proposed, in which only one image is interacted and the other images are segmented automatically with the same foreground target. Firstly, by constructing the feature point locus model, the connection between other image feature points and interactive image feature points is established, then the foreground / background classification of feature points is carried out by using the shortest path algorithm, and the local correspondence between multiple images is obtained. Finally, using the classified feature points as a shape priori, the geodesic convex segmentation algorithm is used to achieve the consistent segmentation of multi-view images. The experimental results of hit rate and false alarm rate show that the algorithm in this paper has a good uniform segmentation ability, and this local correspondence method can also achieve multi-target segmentation.
【作者单位】：南昌航空大学江西省图像处理与模式识别重点实验室;
【基金】：国家自然科学基金(61661036,61462065,61263046) 江西省教育厅科学技术项目(GJJ150738) 江西省图像处理与模式识别重点实验室开放基金(TX201604004)
【分类号】：TP391.41

【相似文献】