Cross-Media Retrieval based on Pseudo-Label Learning and Semantic Consistency Algorithm

To retrieve heterogeneous multimodal data with the same semantics, many algorithms for retrieval over multimodal data have been suggested. The organization and analysis of heterogeneous data have become the focus of intensive research. Here, a new and efficient algorithm for cross-media retrieval is proposed based on pseudo-label learning and semantic consistency (PLSC). In this algorithm, an adaptive learning projection matrix optimization method is proposed, and in the process of learning the projection matrices, the method fully considers the semantic information of the labeled and unlabeled samples. Thus, the PLSC algorithm can utilize more useful information than other methods and can learn the more efficient projection matrices. Firstly, the class centers of labeled text are computed. We use median feature vectors as the class center vectors. Next, unlabeled images are projected onto the text space and are assigned pseudo-labels by comparing with the class center vectors of the text data. Finally, a new training dataset, which includes labeled and unlabeled data, is generated for training the projection matrix. Using the projection matrix to project image or text data onto the same feature space, the data can be compared with each other for similarity, and the distance between data points can be calculated using the Euclidean metric. Validation experiments suggest that the PLSC outperforms other state-of-the-art algorithms.