Gradient-Enhanced Deformation-Aware Defect Detection for Automated Fiber Placement with Synthetic Sample Expansion

Abstract

Existing learning-based AFP inspection methods predominantly rely on grayscale/depth-like representations or directly learned appearance features, while the directional relationship between laser-profile height variation and defect boundaries remains insufficiently exploited. This limitation is particularly relevant to low-contrast and slender AFP defects, for which periodic tow textures can interfere with boundary localization and only limited real defect samples are typically available. To address these issues, this study develops an AFP-specific instance-segmentation framework that coordinates geometry-informed input representation, deformation-aware feature processing, and training-distribution expansion. After outlier correction, RANSAC reference-surface fitting, and residual enhancement, the residual height field and its signed horizontal and vertical gradients are used as a three-channel input to explicitly represent local surface amplitude and directional boundary variation. C2f_DCNv2 and CBAM are incorporated into YOLOv8n-Seg for complementary purposes: adaptive sampling along elongated or irregular defect geometry and suppression of less relevant tow-texture responses, respectively. ConSinGAN-generated samples are screened, manually annotated, and used only for training to expand the represented defect morphology. Three-seed experiments show that the gradient-enhanced representation improves mask P, mask R, mask mAP@50, and mask mAP@50:95 by 2.56, 1.67, 2.85, and 1.22 percentage points over the depth-enhanced input. The architecture modification provides additional gains of 1.32 and 1.52 points in mAP@50 and mAP@50:95, and ConSinGAN augmentation adds 1.19 and 1.51 points. The complete method achieves 94.36% mask P, 91.93% mask R, 94.71% mask mAP@50, and 60.17% mask mAP@50:95. The complete post-acquisition dual-window processing pipeline requires 26.8 ± 0.9 ms per synchronized pair, compared with a 409.6 ms acquisition interval at 30 m/min. These results demonstrate the effectiveness of coordinating AFP-specific geometric representation, morphology-adaptive feature processing, and controlled training-set expansion for online defect segmentation under the evaluated conditions.