In this paper, Hilbert scan based tree representation (HSBT) is presented for image search. Unlike common ways decreasing the number of interest points or reducing the dimensions of features or using searching methods to match interest points, the proposed method builds a tree for each image and gives a new distance measure to calculate the similarity between the query and images in database. In the proposed approach, Hilbert scan for arbitrarily-sized arrays is used to map the interest points from two-dimensional space to one-dimensional space at first. Then, interest points set is divided into several parts by a separation way, and a grouping strategy is given to build a tree for each image. Experimental results show that the proposed approach is space saving. That is because it only stores clustering center and relevant information of each node in the tree. It is also time saving since the similarity calculation is up to the nodes of tree rather than all the descriptors of image. At the same time, the retrieval precision is good, because Hilbert scanning preserves the correlation in two-dimensional image, so nodes of tree are shaped according to the compactness of interest points which can employ the local information as much as possible.