基于LDA图像场景分类方法的增量学习研究

Research on Increment Learning way to LDA Based Image Scene Classification

现有场景分类方法只能识别原训练学习的图像类,对于新增图像类的识别任务,需要将其与原训练类合并后重新训练模型.在LDA(Latent Dirichlet Allocation)的基础上提出一种改进方法来训练生成模型,用于实现自然图像场景分类.根据狄雷克里参数的伪计数作用,改进了LDA模型学习方法.以训练图像的通用主题先验参数作为各类场景主题分布预设先验参数,推导各类场景的类主题构成变化,同时改善了EM参数推导过程中的慢收敛问题,实现了模型增量学习,有效地提高了模型的泛化能力.通过模型计算复杂度比较和增量学习实验对本文方法进行了验证,实验证明本文方法能以较低的时间复杂度取得较高的分类平均正确率.

Prior work discerned scene categories properly, only when they had been trained to learn these training categories. For new unseen categories task, they couldn＇t correctly discern them. Based on LDA （Latent Dirichlet Allocation）, the paper proposed a new way to conquer the problem. Dirichlet parameter did pseudo count for the number of image topic, so it can be used to enforce the LDA model. During the model inference, reasonable category topic prior can be deduced with the basis of general topic composition, which also play a great role to improve the slow convergence problem of EM. Furthermore, increment learning becomes a big light to strength the model＇s generalizing ability. After comparing the iterative times of EM（ Expectation Maximum） among several models, our model showed low computation complication time with best performance. We also investigated the classification performance with classic 13 scene image database. The experiments had demonstrated that our model can get better performance with less training time.