中文摘要：

特征提取是图像分类的关键部分之一。现有的Dense SIFT特征采用固定网格和步长以从上到下、从左到右的重叠方式提取特征,如果图像分辨率过大,将会导致提取的图像特征数量非常大,并且引入大量的冗余信息。为此,提出了一种低冗余Dense SIFT特征提取方法。该方法首先对图像进行预处理,实现对图像的紧凑表示;然后,利用数据中心化思想和e0范数去除冗余的Dense SIFT特征点,节约特征存储所需的空间,降低后续处理的计算复杂度;最后,将低冗余Dense SIFT特征提取方法应用于图像分类,提出了一种图像分类方案。实验结果表明,采用所提出的Dense SIFT特征提取方法,在减少特征点数量的同时,可以提升特征的区分能力。

英文摘要：

Feature extraction is one of the key parts in image classification.The existing Dense SIFT feature method adopts fixed grid and step-size to extract features by scanning way from top to bottom and left to right.If image resolution is too high,more image features will be extracted,so that a lot of redundancy information will be introduced.Therefore,a low-redundancy Dense SIFT feature extraction algorithm is proposed.In this algorithm,the preprocessing is executed on the image,which can produce the compact expression of image.Then,the centralization idea and the e_0 norm are exploited to optimize Dense SIFT features for removing the redundant feature points,in order to finally improve the description ability of features.Finally,the low-redundancy Dense SIFT is applied to image classification.Experimental results show that the proposed scheme can reduce the number of feature descriptors and improve the performance of feature.