Class consistent and joint group sparse representation model for image classification in Internet of Medical Things

Abstract The amount of data handled by Internet of Medical Things (IoMT) devices grows exponentially, which means higher exposure of sensitive data. The security and privacy of the data collected from IoMT devices, either during their transmission to a cloud or while stored in a cloud, are major unresolved matters. Automated human larynx carcinoma (HEp-2) cell classification is critical for medical diagnosis, but most of traditional HEp-2 cell classification algorithms dramatically rely on a single modal feature or fuse different modality features based on fixed weighted schemes, with the result that the complementary information of multimodal features will be not reasonably utilized. In this paper, a class consistent and joint group sparse representation model (CCJGSR) is proposed, expresses the test data through the sparse linear combination of training data and constrains the observations from different modalities of the test object to share their sparse statements. Group sparse representation can fully explore the complementary relationships among different modality features. At the same time, the objective function embeds both the group regularization terms and class consistent, where they enforce the intuitive constraint which the predicted class labels are consistent across all modalities. The experimental results on the HEp2 cell dataset indicate that our proposed algorithm is robust and efficient, and it outperforms existing approaches.