Unsupervised Deep Discriminative Feature Learning for Aerial Image Classification

Nonnegative sparse coding (NSC) is widely utilized as an image representation model. However, conventional NSC methods usually cause large representation errors, lack of spatial information and weak discriminative. In order to overcome these drawbacks, this paper proposes an unsupervised deep discriminative feature learning framework for aerial image classification which is based on Fisher Discriminative Nonnegative Sparse Coding (FDNSC) and Deep Belief Network (DBN). First, image features are extracted by using scale-invariant feature transform (SIFT). Then fisher discriminative analysis is added to construct a NSC with fisher discriminative criterion constraint, thus to obtain the discriminative sparse representation of images. Finally, DBN is combined to perform aerial image classification. The proposed method is applied to OT data set and UC Merced data set. Experimental results show that the proposed method efficiently utilizes spatial information of images and can promote the spatial separability of sparse coefficients, thus improves the classification performance and is more suitable for aerial image classification.