MFDNet: A Lightweight Mobile Mamba-based Detector for Real-Time Aviation Casting Defect Detection

Abstract Detecting defects in aviation castings is critical for ensuring the structural reliability and manufacturing safety of aircrafts. However, traditional manual inspection is time-consuming and prone to human error, whereas existing deep learning-based methods struggle to accurately detect subtle defects, such as microcracks and fine scratches, under resource-constrained industrial conditions. To address these challenges, this paper proposes a Mobile Mamba-based Fusion and Decoupled Detection Network (MFDNet) for aviation casting defect detection. The proposed framework introduces a Mobile Mamba-based Multi-Receptive-Field Feature Interaction backbone(MMB) to enhance long-range contextual modeling of slender defects, a Dynamic Multi-branch Gated Fusion module(DMGF) to adaptively integrate multi-scale features, and a Frequency-aware Decoupled Detection Head(FDDH) to improve defect classification and localization while maintaining efficient real-time performance. Experimental results demonstrate that MFDNet achieves a mean Average Precision (mAP) of 97.4% with an inference speed of 129 FPS while maintaining a compact model size of only 2.9 MB with 1.6 million parameters. Compared with representative lightweight detectors, such as YOLOv8n and YOLOv11n, the proposed MFDNet improves the detection accuracy while significantly reducing the model complexity. Furthermore, MFDNet demonstrated promising cross-dataset adaptability on multiple industrial inspection benchmarks, indicating its potential for deployment in broader resource-constrained inspection scenarios.