基于多尺度先验深度特征的多目标显著性检测方法

显著性检测是近年来国内外计算机视觉领域研究的热点问题，在图像压缩、目标识别与跟踪、场景分类等领域具有广泛的应用.针对大多显著性检测方法只针对单个目标且鲁棒性不强这一问题，本文提出一种基于深度特征的显著性检测方法.首先，在多个尺度上对输入图像进行超像素分割，利用目标先验知识对预显著区域进行提取和优化.然后，采用卷积神经网络提取预选目标区域的深度特征.对高维深度特征进行主成分分析并计算显著性值.最后，提出一种改进的加权多层元胞自动机方法，对多尺度分割显著图进行融合优化，得到最终显著图.在公开标准数据集SED2和HKU_IS的实验表明，与现有经典显著性检测方法相比，本文方法对多显著目标检测更准确.

Multiple Salient Objects Detection Using Multi-scale Prior and Deep Features

Saliency detection has been a hot topic in the field of computer vision in recent years, and has been widely applied in image compression, scene recognition and understanding, target tracking and other applications. Most saliency detection methods are only for a single target and their robustness is not good enough. For this problem, a saliency detection method based on deep feature is proposed. Firstly, the image is over-segmentated with multiple-scales, and the prior knowledge is used to extract and optimize pre-salient regions. Then, the deep features of pre-salient regions are extracted with deep convolution neural networks. The principal component analysis is used to reduce the dimension of deep feature, and the saliency value is calculated. Finally, a weighted multi-layer cellular automata is proposed, and the final saliency map is obtained by fusing the multi-scale segmentation saliency maps with the automata. Experiments on standard datasets SED2 and HKU_IS show that the proposed method is more effective compared with other saliency detection methods.