利用多尺度分析和编组的基于目标的注意计算模型

A Computational Model of Object-Based Attentionusing Multi-scale Analysis and Grouping

模拟生物视觉感知提出一种基于目标的注意计算模型,主要用到两个关键技术:多尺度分析和编组.用于多尺度分析的微分算子从原始图像中提取重要边缘,随后源于格式塔知觉组织规则的轮廓编组过程将边缘组织成感知目标.注意焦点按照各目标显著程度递减的顺序在目标间转移,目标显著程度由边缘重要性、区域对比度和轮廓闭合性共同决定.该模型考虑了目标的独立性和完整性,因此比基于空间的注意有更高的检测精度.多尺度分析为轮廓编组提供了候选边缘,从而提高了编组的效率.对多类自然图像的实验验证了该模型计算上的高效性和生物学上的合理性.

Recent biological experiments have presented increasing evidence for object-based attention. We propose a computational model of object-based attention to simulate biological perception. Two techniques are employed： multi-scale analysis and grouping. Differential geometry descriptor in multi-scale analysis extracted important edges from source images and subsequent contour grouping process organized the edge image into perceptual objects. The later process originated from Gestalt laws. Then focus of attention shifted among objects in order of conspicuousness, which was measured by edge saliency, region contrast and topological property of closure. The proposed model exhibits several advantages. It considers integrality of objects and thus gains higher searching accuracy than space-based attention. It uses multi-scale analysis to select candidates and thus improves efficiency of contour grouping, Experiments on different types of images show high efficiency and biological plausibility of our model.