Abstract:
In order to extract image representation including useful information, a method is proposed based on hierarchical feature learning and label consistent K-SVD. Firstly, a great number of patches are densely sampled only from grey or RGB type of images. Secondly, image features are generated using hierarchical orthogonal matching pursuit instead of traditional pattern based on scale invariant feature transform (SIFT) descriptor combined with spatial pyramid pooling. With a label consistency constraint, a discriminative dictionary is learned by K-SVD algorithm employing the acquired features, as well as an optimal linear classifier. The experiments on Caltech101, Oxford Flowers and UIUC-Sports benchmark datasets show that the proposed method can achieve 76.7%, 84.9% and 87.1% respectively in terms of classification accuracy, which performs better than other algorithms.
