基于多尺度区域对比的显著目标识别  

Salient object detection based on multi-scale region contrast

作　　者：成培瑞 王建立[1] 王斌[1] 李正炜[1] 吴元昊[1] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049

出　　处：《中国光学》2016年第1期97-105,共9页Chinese Optics

基　　金：国家高技术研究发展计划(863计划)资助项目(No.2013AAXXXX042)~~

摘　　要：为了对图像中的显著目标进行更精确的识别,提出一种新的基于多尺度区域对比的视觉显著性计算模型。首先基于多尺度思想将图像分别分割为不同数目的超像素,对超像素内的像素颜色值取平均以生成抽象化图像;然后根据显著特征的稀少性及显著特征的聚集性,计算单一尺度下超像素颜色特征的显著性值;最后通过取各尺度超像素显著度的平均值来融合多尺度显著图,得到最终的视觉显著图。实验表明,以MSRA图库中的1 000张随机自然图片为例,该模型较现有较好的区域对比模型,显著目标识别的精确率提高了14.8%,F-Measure值提高了9.2%。与现有的算法相比,该模型提高了算法对显著目标大小的适应性,减少了背景对显著目标识别的干扰,具有更好的一致性,能更好地识别显著目标。A novel visual saliency computing model is proposed based on multi-scale region contrast to perform more accurate detection on salient object. Firstly, the image is divided into different number of super-pixels based on multi-scale method, and the values of pixels in every super-pixel are averaged to create abstract im- age. Secondly, based on scarcity and aggregation, both of which are the characters of saliency, the color~ sa- liency of super-pixel is computed in single scale. By averaging the salient images in every scale, the multi- scale salient images are fused and the final visual salient image is obtained in the end. The simulation result shows that with 1 000 random nature images in the MSRA Libraries, the model improves the precision ratio of salient object detection by 14. 8% and F-Measure value by 9.2% , compared with current well-performed re- gion contrast model. The model improves the adaptability of the size of salient objects, and reduces the dis-turbance of background. It performs better consistency and has better ability to recognize salient object in com- parison with current algorithms.

关 键 词：显著目标识别 视觉显著性计算模型 多尺度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]