【摘要】 针对BOF模型中的码本训练问题,提出了一种改进的K-means方法。传统的K-means方法没有考虑对采集到的特征进行筛选,基于优化的方法可以看做是一种特征筛选的方法,但是实现复杂,计算量大。提出了一种基于Gist信息的特征筛选方法。根据Gist信息可以将图像粗分为背景区域和前景区域,然后对前景区域进行密集的特征采样,对背景区域进行稀疏的特征采样,最后所获得的特征都用来建立码本。实验结果表明,该方法训练的码本在Caltech101上有很好的分类效果,表明了该方法的有效性。更多还原

【Abstract】 For the dictionary learning problem of bag-of-features(BOF) model,this paper presented an improved K-means method.K-means cannot achieve a good performance for it disregard feature selection and optimized-based method can be regard as a feature selection method,but the computation cost is usually intractable leading to long training time.This paper presented a method to select feature based on Gist information.First,images could be roughly partitioned into foreground and background based on Gist information.Then densely feature extracting was taken on the foreground and sparsely feature extracting was applied on the background.Finally,the method used these features to construct a more discriminative dictionary.Experiments on Caltech101 show that this method can achieve a better performance,and it shows that this method can achieve a better discriminability.更多还原