A lightweight and morphology-aware framework for real-time in-line defect inspection in automated fiber placement

Annals of Forest Science · Springer Nature · 2026-11-13 · DOI: 10.1007/s10845-026-03012-0

Аннотация

Real-time in-line defect inspection is critical for ensuring the manufacturing quality and structural integrity of aerospace composite components produced by automated fiber placement (AFP). However, deploying vision-based inspection systems in practical AFP environments remains challenging due to the highly anisotropic morphology of slender defects, complex and periodic textures of carbon-fiber substrates, and stringent real-time computational requirements. Existing general-purpose object detection models often struggle to capture discriminative features of slender defects while suppressing background texture interference, resulting in degraded detection accuracy under strict inference latency constraints. To address these challenges, this paper proposes AFPDet, a lightweight and high-performance defect detection framework specifically designed for in-line AFP inspection. First, a Frequency–Spatial feature extractor is developed to enhance the continuity of slender defect representations while suppressing high-frequency background interference. Second, an anisotropic–isotropic dual-stream decoupled neck (AIDD) is introduced to reduce feature interference among heterogeneous defect types through direction-aware adaptive sampling. Third, a morphology-decoupled orthogonal regularization (MDOR) loss is designed to impose orthogonality constraints on defect representations with different morphological characteristics, thereby improving feature separability and defect discrimination without introducing additional inference overhead. Extensive experiments on a real-world in-line AFP defect dataset demonstrate that AFPDet achieves an $$\textrm{mAP}_{50}$$ of 85.7% at 105 FPS with only 8.9M parameters. These results demonstrate the effectiveness and practicality of AFPDet for accurate and real-time defect inspection in high-speed AFP manufacturing.

Читать у издателя →

Похожие статьи