Cross-domain segmentation network based on gating feature fusion with decoupled attention

Image segmentation, as a fundamental component of computer vision tasks, has demonstrated significant application value in fields such as medical diagnosis and industrial inspection. Although U-shaped architectures have made significant progress in this field, their limitations in cross-level feature fusion and detail recovery still constrain the segmentation performance in complex scenes. We propose the Attention-Guided Dynamic Gating Network (AGF-UKAN) to address limitations in cross-level feature fusion and detail recovery in U-shaped image segmentation architectures. The framework introduces three synergistic optimizations: an encoder integrating a lightweight convolutional unit with a dual-axis cooperative attention module to enhance local feature discriminability while maintaining parameter efficiency; a decoder employing a hybrid upsampling system combining sub-pixel convolution and channel rearrangement, guided by high-frequency information from the attention module to balance semantic integrity and edge precision; and a dynamic gating mechanism that modulates multiscale feature flows through learnable weights, bridging encoder–decoder semantic gaps. Evaluated on breast ultrasound (BUSI), DeepCrack, and axial defect datasets, AGF-UKAN demonstrates superior segmentation accuracy compared with U-Net variants and state-of-the-art networks. The architecture effectively preserves structural details in complex scenes while optimizing feature fusion efficiency, offering enhanced performance for medical and industrial inspection applications.