DSDNet: A lightweight Divide-and-Conquer Network via deep-shallow feature specialization for deep-sea polymetallic nodule video segmentation

Deep-sea polymetallic nodule video segmentation is essential for seabed resource assessment and underwater perception. However, dense co-occurrence among nodules leads segmentation models to over-rely on collective co-occurrence cues, weakening the representation of intrinsic nodule characteristics. Meanwhile, existing polymetallic nodule segmentation methods typically process consecutive frames independently, resulting in redundant computation and insufficient temporal dependency modeling. To address these challenges, we propose a lightweight Deep-Shallow Divide-and-Conquer Network (DSDNet), which assigns specialized roles to deep and shallow features. For deep features, the Grouped Temporal Dependency Modeling Module (GTDMM) distributes deep feature extraction across consecutive frames. At each time step, GTDMM extracts only one compact group-wise feature and integrates it with cached group-wise features from previous frames, thereby reconstructing a complete high-level representation enriched with temporal information while reducing redundant computation. For shallow features, the Core-Edge Prior-Guided Modulation Module (CPMM) progressively modulates decoder features with core and edge priors in a coarse-to-fine manner, enhancing the representation of intrinsic nodule characteristics under dense co-occurrence. Experiments on a real-world deep-sea polymetallic nodule video dataset collected by the Jiaolong submersible show that DSDNet reaches 0.9573 Dice and 155.56 FPS, demonstrating a favorable balance between segmentation accuracy and inference efficiency. The source code is available at https://github.com/ZIXIANGDAI/DSDNet.git .