特征融合与objectness加强的显著目标检测

摘要

Saliency detection is a fundamental part of computer vision applications, and the goal is to detect important pixels or regions in an image which attracts human visual attention most. By analyzing some recent methods, a new approach is proposed to solve detection errors problems and to enhance the adaptation of features in saliency detection. It detects saliency in the perspective of both object and background and integrates multi features. It extracts color distinctness feature in the perspective of object and extracts boundary prior feature in the perspective of background, and then combines the two features to obtain the corresponding map. In order to keep accuracy, it uses objectness feature to refine the saliency of detected regions. In comparison experiments, it achieves an average precision of 92.4% on MSRA-1000 databases, and achieves higher precision on CSSD dataset and ECSSD dataset. Experimental results demonstrate the used features make up for each other, which can enhance the saliency detection accuracy.%显著目标检测是计算机视觉的重要组成部分，目的是检测图像中最吸引人眼的目标区域。针对显著检测中特征的适应性不足以及当前一些算法出现多检与漏检的问题，提出从“目标在哪儿”与“背景在哪儿”两个角度描述显著性的框架，进行特征融合来提高显著目标检测的准确率。从这两个角度分别提取图像的颜色区别性特征与边界先验特征并进行特征融合，使用objectness特征加强显著性，最终得到显著图。在MSRA-1000数据集上的评估中，该算法达到平均92.4%的准确率，能和最先进算法相媲美；而在CSSD、ECSSD数据集上的实验，该算法有更高的准确率，优势明显。实验结果表明所使用的特征之间能够互相补充，互相弥补“，目标在哪儿”与“背景在哪儿”的检测框架描述图像显著性具有合理性。