MkT-yolo_SKD_P: a lightweight real-time object detector for foreign object retrieval in pressurized water reactors

Due to the challenging conditions of industrial underwater imaging, the variability of target geometries, and stringent computational constraints, real-time foreign object detection in pressurized water reactors poses a significant challenge for nuclear power plant monitoring systems. This paper proposes MkT-yolo_SKD_P, an optimized lightweight detection model, to address these limitations. The model incorporates the MDIC-KAN (Mix Dynamic Inception Conv—Group Rational-Kolmogorov–Arnold Networks) module, which achieves dynamic multi-scale feature fusion through adaptive weighted deformable convolution and combines GR-KAN (Group Rational-Kolmogorov–Arnold Networks) for efficient nonlinear modeling, thereby enhancing feature representation while reducing computational load. Furthermore, the proposed model introduces TDAH (Task Dynamic Alignment Head), a lightweight detection head with task interaction capabilities. It unifies localization and classification tasks through task decomposition and dynamic dependency learning, adaptive spatial alignment using DCNv2 deformable convolution, and the CRCS (Classification Refinement with Context Selection) module. Additional optimization is achieved via BCSKD (Bridging Cross-task Protocol Inconsistency Self-Knowledge Distillation) and LAMP (Layer-Adaptive Magnitude-based Pruning). Experimental results demonstrate that the proposed model outperforms YOLOv11s, achieving higher detection accuracy and significantly improved processing speed while reducing parameter count by 72% and GFlops by 52.6%. When deployed on a GeForce RTX 4070 Laptop, MkT-yolo_SKD_P achieves an FPS (batchsize=8) of 146.32, making it suitable for underwater robotic foreign object retrieval tasks under low-power GPU conditions.