Sparse-Coded Features for Image Retrieval

State-of-the-art image retrieval systems typically represent an image with a bag of low-level features. Since different images often exhibit different kinds of low-level characteristics, it is desirable to represent an image with multiple types of complementary features. The systems scalability is, however, significantly lowered when increasing the number of feature types, as the amount of data is also increased rapidly both in index and in query representation. In this paper, we apply sparse coding to derive a compact yet discriminative image representation from multiple types of features for large-scale image retrieval. We first convert each feature descriptor into a sparse code, and aggregate each type of sparse-coded features into a single vector by max-pooling. Multiple vectors from different types of features are then concatenated and compressed to obtain the final representation. Our approach allows us to add more types of features to improve discriminability without sacrificing scalability. In particular, we design a new micro feature which is complementary to existing local invariant features. By combining our micro feature with various local invariant features using the sparse-coding framework, our final compact representation outperforms the state of the art both in retrieval performance and in scalability.