基于改进流行排序算法的显著性区域检测

Salient region detection based on improved manifold ranking

针对现有基于图的流行排序的显著性检测算法对于复杂背景图像检测中效果不理想问题,提出改进的基于流形排序算法的显著性区域检测.首先将图像分割成4种不同的超像素尺度,并根据图像的RGB,CIELab的颜色特征和LBP纹理特征分别计算4种尺度图像的上、下、左、右4个方向的边界显著图,分别融合不同尺度图像的4个方向的边界显著图得到相应不同尺度图像显著图,融合4种尺度图像的显著图得到弱显著图;然后根据弱显著图以生成强模型的训练样本,通过多核提升算法学习来自输入图像的样本进行强分类,以检测显著像素;最后综合多尺度显著图进一步提高检测性能,并进行优化处理得到最终的显著图.为了验证该算法的正确性和有效性,在公开数据集MSRA1000、ECSSD和PASCAL-S上进行仿真实验,实验结果表明,该算法不仅能够得到较好的视觉效果,而且召回率、准确率和F-measure等评价指标比传统算法有明显提升.

Existing graph-based manifold ranking based salient object detection saliency detection algorithm is less effective in detecting images with complex background.This paper proposes an improved saliency detection algorithm based on manifold ranking.Firstly,the image is divided into four different super-pixel scales.Then according to the RGB color space feature,the CIELab color space feature and the LBP features,calculated the boundary saliency maps of the upper,lower,left and right directions of the four scale images.The saliency maps of four directions are combined to obtain the four-scale image saliency map,respec-tively.The weak saliency map is computed by the saliency map of the four-scale image.Secondly,a training sample generated based on the weak saliency map.A strong classifier based on samples directly from an input image is learned to detect salient pixels by the multiple kernel boosting algorithm.Finally,the integrated multi-scale saliency map further improves the detection performance.In addition,the process is optimized to get the final saliency map.Furthermore,the optimization is carried out for better performance.The experiments on public data sets MSRA 1000,PASCAL-S and ECSSD show that our algorithm not only achieves good vision effect,but also improve the performance evaluation of precision-recall curves and F-measure values.