An improved method of book cover retrieval

SIFT is a well-known feature extract algorithm for image matching. Due to the high dimensionality caused by the feature descriptor, it is too time consuming for real-time image retrieval application with kd-tree. To fast the image retrieval process, we deal with it in two aspects. Firstly, a novel method is proposed to reduce the dimensionality of the sift descriptor, which significantly drops the memory consumption caused by the millions of key points. Secondly, in kd-tree matching work, an improved decision algorithm is used for skipping some Euclidean distance computation. Experiment on our large book cover dataset demonstrated that the keypoint matching time can be reduced by over 25%.