A Deep Learning‐Assisted Non‐Contact Measurement Method for Extracting Formed Arc Radius in Continuous Fiber‐Reinforced Composite Additive Manufacturing

ABSTRACT Accurate and efficient measurement of filament bending trajectories remains a significant challenge in complex‐path additive manufacturing of continuous fiber‐reinforced polymer (CFRP) composites. To address this issue, this study proposes a deep learning‐assisted non‐contact approach for accurately extracting geometric parameters of curved filaments in CFRP additive manufacturing. A semantic segmentation model based on a U‐Net architecture with a convolutional block attention module (CBAM) and Visual Geometry Group 16‐layer network (VGG16) backbone—referred to as VC‐UNet—is employed to tackle challenges such as complex curvature and blurred boundaries. To extract bending geometry with high fidelity, an adaptive arc detection pipeline is developed, integrating line‐based region masking with an adaptive Hough transform with local thresholding (AHT‐LT). The resulting arc features are further refined using least‐squares circle fitting and mapped to physical scale via calibrated camera parameters. Comparative analysis with conventional ImageJ measurements demonstrates strong consistency, with over 95% of the predicted radii falling within the Bland–Altman limits of agreement and a Pearson correlation coefficient of 0.9909. These results validate the method's accuracy, repeatability, and applicability for online defect monitoring and process feedback in advanced manufacturing environments.