A Comparative Study On Features for Similar Image Search

Feature representation plays a key role to the success of an image retrieval system. In this paper, a comparative study over the effectiveness of several features for content-based image search is presented. This study covers across several conventional features as well as convolutional neural networks (CNN) features, which are introduced recently into retrieval tasks. In particular, the evaluation is conducted when features are under the same encoding scheme. In addition, a hybrid feature representation that combines key-point detector and CNN descriptor is proposed, in which the geometric invariances of keypoint feature and the distinctiveness of CNN feature are integrated. Experiments on popular evaluation benchmarks show that this hybrid feature achieves superior performance.