非监督学习图像层次组合模型的研究算法

Unsupervised Learning of Hierarchical Compositional Models in Image

针对传统的监督学习图像模型对训练样本要求苛刻的问题,本文提出一种非监督学习算法,该算法不仅对训练样本要求简单,而且学习到的层次组合模型由能在位置和方向进行扰动Gabor小波组成,是一种可变形模板,因此一定程度上提高定位及分割算法在物体发生形变情况下的鲁棒性。经过多组实验结果表明,本文所提出的层次组合模型能高效地解决目标在发生形变、存在遮挡以及复杂背景下的定位分割问题。

Because the traditional supervised learning algorithm is rigid to training images, this paper proposed an unsupervised method for learning hierarchical compositional models for representing natural images. The method is very simple to training images,and each model is in turn a composition of Gabor wavelets that are allowed to shift their locations and orientations, so it is robust when target occurred a little deformation or the presence of occlusion. The experimental results show that hierarchical compositional models can solve the problem of localization and segmentation when the target is partly changed, occlusive or in complex background.