D-SIFT：一种面向DCT域的扩展SIFT特征

摘要

提出一种面向DCT域的扩展SIFT特征(即D-SIFT特征)，以提高JPEG图像的匹配速度。 D-SIFT特征首先在 DCT域对图像进行不同尺度的高斯滤波，构建出图像的尺度空间；然后，把滤波后的图像划分为大小为2×2的图像块，并通过检测出极值图像块来估计关键特征点位置；最后，利用当前块周围16×16个子块的 DCT直流系数来统计主方向，提取每个极值块的描述子。实验结果表明， D-SIFT特征可以有效地提高JPEG图像的匹配效率。%To accelerate the speed of JPEG image matching, we propose a novel local feature called D-SIFT feature, which is an extension of SIFT in the DCT domain. Firstly, D-SIFT does Gaussian filtering to the image in the DCT domain, obtaining the image scale space. Then, D-SIFT divides all the filtered images into 2×2 image blocks, and estimates all the key points by detecting the key blocks. Finally, D-SIFT obtains the main orientation and extracts the descriptor for each key block from its surrounding 16×16 blocks by using their DC coefficients. Experimental results show that, D-SIFT feature can efficiently improve the speed of JPEG image matching.