摘要
针对已有视觉注意模型在整合对象特征方面的不足,提出一种新的结合高层对象特征和低层像素特征的视觉注意方法。首先,利用已训练的卷积神经网(CNN)对多类目标的强大理解能力,获取待处理图像中对象的高层次特征图;然后结合实际的眼动跟踪数据,训练多个对象特征图的加权系数,给出对象级突出图;紧接着提取像素级突出图,并和对象级突出图融合获得显著图;最后,在OSIE和MIT数据集上验证了该方法,并与国际上流行的视觉注意方法进行对比,结果显示该算法在OSIE数据集上获得的AUC值相对更高。实验结果表明,所提方法能够更加充分地利用图像中对象信息,提高显著性预测的准确率。
Concerning the defects of fusing object information by existing visual attention models, a new visual attention method combining high-level object features and low-level pixel features was proposed. Firstly, high-level feature maps were obtained by using Convolutional Neural Network （CNN） which has strong understanding of multi-class targets. Then all object feature maps were combined by training the weights with eye fixation data. Then the saliency map was obtained by fusing pixellevel conspicuity map and object-level conspicuity map. Finally, the proposed method was compared with many popular visual attention methods on OSIE and MIT datasets. Compared with the contrast methods, the Area Under Curve （AUC） result of the proposed method is increased. Experimental results show that the proposed method can make full use of the object information in the image, and increases the saliency prediction accuracy.
出处
《计算机应用》
CSCD
北大核心
2016年第11期3217-3221,3228,共6页
journal of Computer Applications
基金
华为创新基金资助项目(YJCB2010022IN)~~
关键词
视觉注意
自顶向下
显著性
对象信息
卷积神经网
visual attention
top-down
saliency
object information
Convolutional Neural Network （CNN）
