Abstract:

The traditional bag-of-features (BoF) model for object recognition assumes each local feature point is related to only one visual word. Besides, sparse coding with l₁-norm constraint generally selects only one feature without concern for which one is selected. A novel bag-of-features model based on elastic net sparse coding is presented in this paper. The model uses scale invariant feature transform (SIFT) feature descriptors to construct a feature dictionary, and then applies an elastic net regression model to the solution of sparse-coefficient vectors. Finally the sparse-coefficient vectors in each object image are pooled for classification. Compared with the conventional BoF model and the BoF model based on l₁-norm sparse coding, our model achieves better recognition performance and is more robust to the variation of viewpoints. Experiments on the space object image database demonstrate the effectiveness of the proposed model.

Key words: bag-of-features, l₁-norm constraint, sparse coding, elastic net regression, space object recognition, feature extraction