基于边界扩展的图像显著区域检测
Image salient region detection based on boundary expansion

在显著区域检测中，背景先验已被证明有效。通常，图像的边界图像块被假设为背景，其他图像块根据与边界图像块之间的差异来确定显著性，差异越大则显著性越强。然而，当图像背景杂乱或者前景与图像边界有重叠时，仅仅利用边界图像块作为背景将会产生包含较强噪声的显著图，从而使得检测精度下降。该文首先将图像边界图像块向图像内部扩展，使其包含尽可能多的背景像素；然后，利用未扩展到的图像块作为前景查询项，采用二级排序算法来度量所有图像块的显著性。在3个公开的复杂显著区域检测数据集上的大量实验表明该算法优于其他5种算法。
Abstract：Background priors have been shown to improve salient region detection. Typically, image boundary patches are assumed to be the background and the saliency of other patches is defined by their difference from the boundaries. A greater difference indicates a more salient patch. However, when the background is cluttered, or the foreground overlaps the image boundary, using only boundary patches to indicate the background may lead to a saliency map with strong noise and compromise the detection accuracy. To address this problem, the boundary patches are first expanded here into the image interior to contain as much background as possible. Then, the rest of the patches are used as foreground queries with the saliency of each patch measured by a two-stage ranking algorithm. Tests on three large public datasets demonstrate the superiority of this method over five other algorithms.