MFAFNet: Multiscale frequency-adaptive fusion network for domain generalized underwater object detection

Underwater object detection (UOD) in unknown underwater environments remains challenging due to domain-dependent image degradation, which causes unstable feature representations and substantially reduces cross-domain generalization. To address this problem, we propose MFAFNet, an end-to-end Multiscale Frequency-Adaptive Fusion Network for domain-generalized underwater object detection. MFAFNet consists of three complementary components that progressively enhance, extract, and fuse degradation-robust features. First, the Frequency Adaptive Enhancement Network (FAENet) decomposes the input image into low- and high-frequency components and adaptively enhances global structural information and local edge and texture details affected by underwater degradation. Second, the Faster Cross Gated Aggregation block (FCGA) combines partial convolution with cross-gated spatial-channel interaction to efficiently extract robust feature representations while maintaining low computational complexity. Finally, the Multiscale Adaptive Fusion Hierarchical Network (MAFNet) adaptively integrates semantic and detailed information across multiple feature scales, improving object discrimination under domain shifts. Extensive experiments on multiple underwater datasets demonstrate that MFAFNet achieves robust detection performance and strong cross-domain generalization.