Boosted multifold sparse representation with application to ILD classification

Classification performance with sparse representation is largely affected by the low discriminative power of image features. In this study, we propose a new sparse representation model, namely the Boosted Multifold Sparse Representation (BMSR), to improve the classification performance. By dividing the training set into multiple subsets, sparse representation using one subset is used as a weak classifier. A threefold boosting approach is then designed to combine the multiple weak classifiers to create the final class label. We applied the BMSR method to classify image patches of different interstitial lung disease (ILD) patterns using a publicly available dataset. Promising performance improvement over non-boosted sparse representation is shown.