改进背景先验和前景先验的显著性检测

针对传统的流行排序显著性检测算法存在的问题,本文提出了改进背景先验和前景先验的显著性检测.首先计算图像的凸包,并将图像分割成不同尺度的超像素;然后以凸包区域之外的超像素为背景种子,结合多尺度下图像的多种底层特征得到最终的背景显著图;第三,以凸包区域之内的超像素为前景种子,结合多尺度下图像的多种底层特征得到最终的前景显著图;第四,融合最终的背景显著图和最终的前景显著图得到弱显著图,通过多核增强(MKB)算法对由弱显著图生成的训练样本进行强分类,生成强显著图;最后综合强弱显著图,得到最终的显著图.通过在MSRA1000,PASCAL和ECSSD数据集上与其他13种算法进行对比,验证了本文算法在显著目标检测的准确性方面更具优势.

Saliency Detection Based on Improved Background Prior and Foreground Prior

Aiming at the problems of the traditional manifold ranking saliency detection algorithms,this paper proposes an improved background prior and foreground prior saliency detection.First,the convex hull of the image is calculated,and the image is divided into superpixels of different scales.Then the superpixels outside the convex hull area are used as background seeds,and the various low level features of the image at multiple scales are combined to obtain the final background saliency map.Third,the superpixels in the convex hull region are used as the foreground seeds,and the final foreground saliency map is obtained by combining various low level features of the image at multiple scales.Fourth,the final background saliency map and the final foreground saliency map are fused to get the weak saliency map,and the training samples generated by the weak saliency map are strongly classified by the multi-core enhancement(MKB)algorithm to generate the strong saliency map;finally,the strong and weak saliency map are synthesized to get the final saliency map.Through three types of public data sets MSRA1000,PASCAL and ECSSD,compared with the current 13 state-of-the-art algorithms in terms of accuracy,the algorithm performs well.