GKSH: Graph based image retrieval using supervised kernel hashing

The explosive growth of the massive image database brings great challenge to the fast and accurate image retrieval. To address this issue, we propose a kind of supervised hashing method based on the graph representation of the images, which translates images into representative attribute structural graphs. Compared with the traditional supervised methods, this kind of structural graphs can consider the spatial relations among regions in the image, and integrate unsupervised properties of images and supervised information of labels by adopting hashing with graph kernel based on random walks. The learnt hash codes can be a good tradeoff among retrieval speed, memory requirements and retrieval accuracy. The experiments on three image datasets PASCAL, MNIST and HOLIDAY demonstrate that, with the hash codes of the same length, the proposed supervised hashing method can achieve a higher precision and a better efficiency.