基于潜在Dirichlet分布的图像分层表示模型

摘要

The existing image hierarchical representation methods are strict in feed-forward style, and therefore it is not able to solve problems like local ambiguities well. In this paper, a probabilistic model is proposed to learn and deduce all layers of the hierarchy together. Specifically, a recursive probabilistic decomposition process is taken into account, and a generative model based on latent Dirichlet allocation with pyramidal multilayer structure is derived. Two important properties of the proposed probabilistic model are demonstrated:adding an additional representation layer to improve the performance of the flat model and adopting a full Bayesian approach which is better than a feed-forward implementation of the model. Experimental results on a standard recognition dataset show that the proposed method outperforms the existing hierarchical approaches, and it improves the classification and the learning accuracy with better performance.%现有的图像分层表示方法严格局限于前馈型方式，不能较好地解决局部模糊性等问题。基于此，文中提出一种学习和推断层次结构所有分层的概率模型，它考虑递归的概率分解过程，通过推导得到金字塔式多层结构的潜在Dirichlet分布的衍生模型。该模型存在两个重要特性：增加表示层可提高平面模型的性能；采用全Bayesian概率方法优于其前馈型实现形式。在标准识别数据集上的实验结果表明，与现有的分层表示方法相比，该模型表现出较好性能。