Image Classification via Sparse Representation and Subspace Alignment

Image representation is a crucial problem in image processing where there exist many low-level representations of image, i.e., SIFT, HOG and so on. But there is a missing link across low-level and high-level semantic representations. In fact, traditional machine learning approaches, e.g., non-negative matrix factorization, sparse representation and principle component analysis are employed to describe the hidden semantic information in images, where they assume that the training and test sets are from same distribution. However, due to the considerable difference across the source and target domains result in environmental or device parameters, the traditional machine learning algorithms may fail. Transfer learning is a promising solution to deal with above problem, where the source and target data obey from different distributions. For enhancing the performance of model, transfer learning sends the knowledge from the source to target domain. Transfer learning benefits from sample reweighting of source data or feature projection of domains to reduce the divergence across domains. Sparse coding joint with transfer learning has received more attention in many research fields, such as signal processing and machine learning where it makes the representation more concise and easier to manipulate. Moreover, sparse coding facilitates an efficient content-based image indexing and retrieval. In this paper, we propose image classification via Sparse Representation and Subspace Alignment (SRSA) to deal with distribution mismatch across domains in low-level image representation. Our approach is a novel image optimization algorithm based on the combination of instance-based and feature-based techniques. Under this framework, we reweight the source samples that are relevant to target samples using sparse representation. Then, we map the source and target data into their respective and independent subspaces. Moreover, we align the mapped subspaces to reduce the distribution mismatch across domains. The proposed approach is evaluated on various visual benchmark datasets with 14 experiments. Comprehensive experiments demonstrate that SRSA outperforms other latest machine learning and domain adaptation methods with significant difference.