A general image fusion framework using multi-task semi-supervised learning

• Published in: Information fusion Vol. 108; p. 102414
• Main Authors: Wang, Wu, Deng, Liang-Jian, Vivone, Gemine
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2024
• Subjects: Fusion rule; Image fusion; Laplacian pyramid; Medical images; Multi-task; Remote sensing; Semi-supervised learning; Laplacian pyramid; Multi-task; Semi-supervised learning; Medical images; Fusion rule; Image fusion; Remote sensing
• ISSN: 1566-2535; 1872-6305
• DOI: 10.1016/j.inffus.2024.102414

Existing image fusion methods primarily focus on solving single-task fusion problems, overlooking the potential information complementarity among multiple fusion tasks. Additionally, there has been no prior research in the field of image fusion that explores the mixed training of labeled and unlabeled data for different fusion tasks. To address these gaps, this paper introduces a novel multi-task semi-supervised learning approach to construct a general image fusion framework. This framework not only facilitates collaborative training for multiple fusion tasks, thereby achieving effective information complementarity among datasets from different fusion tasks, but also promotes the (unsupervised) learning of unlabeled data via the (supervised) learning of labeled data. Regarding the specific network module, we propose a so-called pseudo-siamese Laplacian pyramid transformer (PSLPT), which can effectively distinguish information at different frequencies in source images and discriminatively fuse features from distinct frequencies. More specifically, we take datasets of four typical image fusion tasks into the same PSLPT for weight updates, yielding the final general fusion model. Extensive experiments demonstrate that the obtained general fusion model exhibits promising outcomes for all four image fusion tasks, both visually and quantitatively. Moreover, comprehensive ablation and discussion experiments corroborate the effectiveness of the proposed method. The code is available at https://github.com/wwhappylife/A-general-image-fusion-framework-using-multi-task-semi-supervised-learning. [Display omitted] •Our method can solve several image fusion problems with a single model.•We propose a multi-task semi-supervised training method to extract complementary information from different tasks.•We propose the PSLPT, which decomposes the source images into features of different frequencies and adaptively learns to merge these features.