DFSC: Infrared-Visible Image Fusion Based on Dual Driving of Feature Optimisation and Semantic Coupling

Infrared-visible image fusion technology is crucial for the preprocessing of advanced visual tasks and has found extensive applications in areas such as all-weather surveillance, military security, medical imaging, and fire monitoring, among others. However, existing fusion methods are insufficient in deeply exploring the essential features of images and establishing robust semantic relationships. Additionally, these methods often struggle to meet the efficiency and real-time performance requirements when processing large-scale image data in practical applications. To address these limitations, this study introduces a novel infrared-visible image fusion approach based on feature optimization and semantic coupling. This approach integrates feature processing with semantic analysis, resulting in significant improvements in both fusion quality and efficiency. A dual-branch encoder network architecture is proposed, leveraging the strengths of Convolutional Neural Networks (CNNs) and Transformers to achieve high-precision optimization of the feature extraction process. Furthermore, a semantic coupling mechanism is incorporated to effectively mitigate information loss, enhancing the overall quality of the fused images. This study takes full advantage of the lossless transmission properties of Invertible Neural Networks (INN) and the long-range dependency modeling capability of Transformers. By coordinating the decoder with a dual-attention mechanism, a two-phase training strategy and an adapted loss function are implemented to ensure efficient feature processing and deep fusion, thus enabling high-efficiency, real-time fusion. Experimental results demonstrate that the proposed method yields superior fusion performance on the TNO and MSRS datasets. In target detection tasks using the M3FD dataset, the method shows exceptional performance, offering an innovative and effective solution for multimodal image fusion. Moreover, the method significantly improves adaptability in complex scenes, overcoming the inherent limitations of traditional fusion methods. This approach is expected to have a considerable impact on the advancement of related fields, providing substantial theoretical and technological support for further technological innovations and application expansions.