Abstract: Because Scale Invariant Feature Transform(SIFT) descriptor is likely influenced by illumination changes, this paper proposes a kind of Discrete Cosine Transform(DCT)-based local invariant feature descriptor. The descriptor uses the characteristics of DCT, ignoring high-frequency coefficients, and reduces the dimension. It is composed of a small numbers of compositions of low-frequency coefficient matrix. As the sign of the DCT frequency coefficient is not sensitive to illumination changes, the proposed descriptor improves the descriptor distinction by setting penalty factor in the calculation of the distance between the descriptors. Result of the implementation test shows that the proposed descriptor has better significance, recall rate and precision than SIFT descriptor.

Key words: Scale Invariant Feature Transform(SIFT), local invariant feature, Discrete Cosine Transform(DCT), image descriptor, image match, penalty factor

摘要： 针对尺度不变特征变换(SIFT)描述子受光照变化影响较大的缺点，提出一种基于离散余弦变换(DCT)的图像局部不变特征描述子。在DCT变换的基础上，忽略高频系数，使用少数中低频系数组成特征矩阵，以降低描述子的维数。利用DCT频率系数正负性对光照变化不敏感的特点，在计算描述子间距离时设置惩罚因子，以提高描述子的可区分性。测试结果表明，与SIFT描述子相比，该描述子具有较好的显著性，且查全率和查准率较高。

关键词: 尺度不变特征变换, 局部不变特征, 离散余弦变换, 图像描述子, 图像匹配, 惩罚因子

CLC Number: 

• TP911.73