Unsupervised feature learning with C-SVDDNet

•A single layer network—CSVDDNet is proposed for unsupervised feature learning.•Unsupervised feature learning methods can be useful when training set is small.•Networks with different receptive field can be combined to make a better prediction.•SIFT representation can be used in unsupervised feature learning network.

In this paper we present a novel unsupervised feature learning network named C-SVDDNet, a single-layer K-means-based network towards compact and robust feature representation. Our contributions are three folds: (1) we introduce C-SVDD encoding, a generalization of the K-means local encoding that adapts to the distribution information and improves the robustness against outliers; (2) we propose a method that effectively embeds the spatial information of 2D data into the final representation based on a modified SIFT descriptor; and (3) we extend our C-SVDDNet to exploit multi-scale information for better feature learning. Extensive experiments on several popular object recognition benchmarks, such as STL-10, MINST, Holiday and Copydays shows that the proposed method yields comparable or better performance than that of the previous state-of-the-art unsupervised feature learning methods.