基于连通图的视觉显著区域检测研究

Visual saliency detection based on connected graph

显著性区域检测是指自动识别出图像中最感兴趣、最重要的区域,目前在目标识别、图像检索等领域应用广泛。基于图的流形排序的显著区域检测算法虽然能够准确高效地检测出一幅图像中的显著性区域,但该算法中使用的K正则图描述的各顶点的空间连接性的图的结构存在局限。为解决上述局限性,研究构造一个更一般的连通图,在显著目标较大或显著目标不连续的情况下,能够更准确地检测出显著性区域。通过在CSSD、SOD、ASD和SED2四个标准数据集上进行大量验证性实验,与六种现有的代表性方法相比,实验结果在PR曲线、F值、MAE等多个指标均表明改进算法有明显的提高,有效验证了算法的有效性。

Visual saliency detection is automatically to detect the most interesting and important area, which is widely used to many applications, such as object recognition, image retrieval and so on. The graph based manifold ranking method could detect the salient area effectively, but the constructed K-regular graph could not represent the image completely. In order to solve the above -mentioned limitation, it constructed a new connected graph to indicate the relationship between image nodes, which could get more accurate results especially on images who had large or discontinuous saliency map. It performed extensive experiments on four benchmarks comparing with six state-of-the-art methods. The results of this method are better than other methods by the standard of the PR-curve, MAE and so on.