摘要
该方法利用视觉底层特征颜色、方向、强度和轮廓构建显著图,通过最大熵估计方法获得显著特征分割蒙板;利用中层视觉特征对图像进行超分割,其中在聚类时考虑特征向量的空间信息,并依据显著性自动分配初始参数,使分割后的超像素与目标轮廓更接近;最后将底层视觉特征和中层视觉特征融合,通过底层特征分割蒙板判定图像的超像素归类,将不规则目标从背景中分离。实验结果表明:本文分割方法受复杂背景和光照的影响较小,分割目标轮廓准确,实现了不规则显著目标与复杂背景的有效分离。
A salient object segmentation method based on low-level visual feature and middle-level visual cues was proposed. First, the low-level visual feature of the original image was extract via color, intensity, orientation and local energy feature channels to build the saliency map. The salient feature mask was acquired via the maximum information entropy principle. Then In middle-level, the visual cues were applied for over-segmentation of an image into superpixels. In clustering, the spatial information of the feature vector was taken into consideration according to the salient intensity, and the initial parameters were automatically set. Thus, the superpixels after segmentation accurately approach the object contour. Finally, for segmenting the irregular object from background, the superpixels were classified using the salient feature mask, and the low-level and middle-level features were fused. The experiment results demonstrate that the proposed method is less sensitive to complex illumination and background, and can be used to segment contour accurately. Moreover, it can be applied to segment irregular objects from complex background.
出处
《吉林大学学报(工学版)》
EI
CAS
CSCD
北大核心
2014年第4期1140-1144,共5页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目(61101155)
吉林省科技发展计划项目(20140101184JC)
关键词
计算机应用
超像素
图像分割
视觉显著特征
显著图
computer application
superpixels
image segmentation
visual salient features
saliency map