AFMF‑Net: A lightweight adaptive feature modulation and fusion network for infrared image super‑resolution

Compared to natural images, infrared images often suffer from lower resolution, reduced contrast, and sparse high-frequency details due to the physical limitations of imaging sensors. Image super-resolution (SR) techniques play a crucial role in enhancing structural details and restoring texture information in infrared imagery. However, it remains a challenging task to accurately model both edge and background features in infrared images while maintaining a lightweight network architecture. To address this issue, we propose a lightweight and efficient infrared image SR network, termed Adaptive Feature Modulation and Fusion Network (AFMF‑Net). The network significantly improves infrared image reconstruction through two core modules while maintaining model efficiency. The first module, the Adaptive Feature Modulation Module (AFMM), integrates dynamic convolution with the Multi-Kernel Estimation Module (MKEM). The MKEM generates multiple independent convolution kernels for each spatial position and applies them to modulate the initial features in a position-specific manner. This design substantially enhances the network's sensitivity to edge structures. The second module, the Nested Dual-Branch Feature Fusion Module (NDFFM), establishes bidirectional interactions along spatial and channel dimensions. This configuration enables effective integration of edge and background features and further improves overall reconstruction quality. Compared with existing methods, NDFFM offers distinct advantages in multi-scale feature fusion and detail preservation. Extensive experiments conducted on multiple standard infrared image datasets demonstrate that our method outperforms existing state-of-the-art approaches in both subjective visual quality and objective reconstruction metrics, exhibiting strong generalization ability and promising application potential.