摘要
针对现有显著性分割算法在图像背景复杂时先验知识不够健壮的问题,提出一种融合注视点预测和流形学习的显著性目标分割算法,能有效地对复杂场景中的显著性目标进行分割.该算法通过引入注视点先验知识和提取超像素分割图,预测并粗分割场景中的显著性目标;为了进一步提高显著性分割的性能,利用色彩模型(CIE-Lab)空间的颜色对比度表示超像素的特征;通过基于流形学习的方法对粗分割区域进行显著性优化,提高了分割精度.实验结果表明:在处理复杂图像集过程中,相比其他分割算法,该算法性能提高了21.8%,并且在不同环境下的显著性目标分割的鲁棒性更好.
As the priors of existing saliency segmentation methods are not robust enough in the complex background, an algorithm which merged fixation prediction and manifold learning was proposed to effectively segment salient objects in complex scenes. The algorithm predicted and segmented salient objects in scenes by introducing the prior of fixation and extracting the map of superpixels. To further improve the performance of saliency segmentation, the algorithm leveraged color contrast be- tween superpixels as features in CIE-Lab （color model） space and resolved the saliency optimization of coarse regions via a manifold learning-based method which improved the segmentation accuracy. Experimental results show that the proposed method has an improvement of 21.8%o than the other best methods on complex datasets and is more robust to segment salient objects in different environments.
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2016年第10期64-69,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(71301057)
上海航天科技创新基金资助项目(SAST201409)
关键词
显著性目标分割
注视点预测
流形学习
色彩模型
超像素分割
salient object segmentation
fixation prediction
manifold learning
CIE-Lab
superpixel segmentation
