从复杂的自然图像中获取目标轮廓是计算机视觉中的经典难题,而提供符合人类感知特性和自然图像统计规律的线索合并模型是提高轮廓质量的关键问题。利用连续性和相似性线索进行轮廓编组,提出一种线索合并模型,拟合格式塔规则中连续性和相似性的统计联合条件概率。该线索合并模型解释了如何用两个相互独立的线索变量得到两个相关线索联合分布的特殊形式,克服了判别式模型刻意回避的相关线索合并问题,是更符合自然图像统计特性和人类感知特性的格式塔线索量化模型。将该模型应用于自然图像的轮廓提取中,实验结果证实了模型的有效性。

Using contour grouping to get perceptual objects from natural images is a traditional difficult problem in computer vision. To provide a cue combination model, which accords with human visual perception and the statistical of the natural images, is the key problem of the definition of grouping cues and the improvement of grouping quality. According to the conditions when a linear cue combination rule works, a cue combination model for contour grouping, which fits the ecological joint distributions of continuity and similarity in Gestalt cues, is proposed. The cue combination model helps to explain how two independent cue channels could give rise to the particular form of the joint distribution of two correlated cues. It also overcomes the cue combination problem between correlated cues which can't be resolved by discriminative model. The proposed model is a quantitative model of Gestalt cues which accords with the statistical of the natural images and human visual perception better.