基于树型贝叶斯网络的场景分类引擎训练算法

Tree structured Bayesian network learning algorithm for scene classification

贝叶斯网络在场景分类统计模型设计中得到广泛的应用。但现有的大部分贝叶斯网络场景分类引擎没有能够充分利用贝叶斯网络丰富的知识表现能力和有效的自动学习能力。首先提出了一种灵活的树型贝叶斯网络分类引擎,用于场景分类模型的设计。然后,以条件对数似然评价为标准研究这种模型的自动学习方法,通过对分类器等价类的研究,证明了树型贝叶斯网络分类引擎自动训练过程可以忽略网络中边的方向,并提出了一个不需要对边重定向的学习算法。由于通常的场景图像编码维度较高,省略了边的重定向过程能够有效地减少模型的训练时间。实验结果验证了所提算法的平均训练时间在基准场景图像库上比传统算法的减少23.32%。

Bayesian network is an effective knowledge representation and inference engine for scene classification.However,most of current researches in semantic image understanding are limited to nave Bayesian network with strong assumptions of independence among features.To address this problem,we formalize a tree structured Bayesian network and introduce it to scene classification.In order to reduce the training time of a tree structured Bayesian network,we demonstrate that it is useless to pay attention to the directions of arcs in a tree structured Bayesian network classifier using log conditional likelihood criterion and propose an algorithm without taking arc reversal into account.Experimental results show that the average training time of our algorithm is reduced by 23.32% compared with that of traditional algorithm.