Feature point matching is a key base of many computer vision problems. A large advantage of fern algorithm, simple and fast, is noticed among the existing algorithms, but classifier trained by fern is too large to low memory device, such as cellphones. For cutting down the size of a classifier, this paper proposes an improved version of fern, named Oriented Fern(OFern), which“normal”patches are done that training for a classifier, and features are extracted, a Naive Bayesian model is built to train a classifier. Experimental results show that compared with the traditional fern, OFern can save memory to 1/8~1/16 at the similar recognition rate, while it still keeps high speed for real-time applications.%特征点匹配是计算机视觉领域研究的核心问题之一。现有的随机蕨算法具有简单、高速的优点，但随机蕨算法训练得到的分类器体积过大，低内存的移动设备难以承受，严重限制了该算法的应用范围。针对该问题，提出一种基于方向信息的随机蕨特征匹配算法，对用于训练的小图块进行“归零化”处理，提取特征属性构造特征向量，建立朴素贝叶斯模型训练分类器。实验结果表明，经过该方法处理后，在相近识别精度下，得到的分类器体积减小到原始算法的1/8~1/16，满足实时性要求。
展开▼
