MAF-Net: A Lightweight Morphology-Adaptive Feature Fusion Network for Semantic Segmentation

Semantic segmentation enables pixel-level scene understanding, and deploying it on resource-constrained platforms such as unmanned aerial vehicles and embedded sensing devices has become increasingly important. In complex scenes, however, objects with substantial scale variation and heterogeneous morphology often lead to missed small objects, discontinuities along slender structures, and incomplete boundaries. To address the challenges, this paper proposes a lightweight semantic segmentation network, named MAF-Net. The network adopts MobileNet as the backbone and enhances the representation of objects with different scales and morphological characteristics through multi-scale feature extraction and top-down feature fusion, while maintaining low model complexity. To further accommodate variations in object scale and structure, a Morphology-Adaptive Kernel Decoder (MAKD) is designed. The module introduces large-kernel, small-kernel, horizontal-strip, and vertical-strip convolution branches, and adaptively fuses the multi-branch features through a gating mechanism, thereby providing complementary contextual representations with different receptive-field scales and directional sensitivities. In addition, a Triple Feature Fusion Refinement Module (TFFRM) is constructed to further refine the initial segmentation results by integrating shallow spatial details, learned boundary-aware responses, and high-level semantic features, thereby alleviating problems such as discontinuities in slender structures and blurred object boundaries. Experimental results show that the proposed method achieves favorable overall segmentation performance on the COCO2017-derived ball-sports dataset, UDD6, and WHDLD, while maintaining a relatively small parameter count and high measured inference speed.