Real-Time Polarimetric Imaging and Enhanced Deep Learning Model for Automated Defect Detection of Specular Additive Manufacturing Surfaces

Additive manufacturing (AM) technology has found extensive applications in aerospace, medical, and automotive fields. Defect detection technology remains a research focus in AM process monitoring. While machine learning and neural network algorithms have recently achieved significant advancements in innovative applications for AM defect detection, practical implementations still face challenges, including insufficient detection accuracy and poor system robustness. To address these limitations, this study proposes the YOLOv5-CAD defect detection model. Firstly, the convolutional block attention module (CBAM) is introduced into the core feature extraction module C3 of the backbone network to enhance attention to critical information and improve multi-scale defect target adaptability. Secondly, the original CIoU loss function is replaced with the Alpha-IoU loss function to accelerate network convergence and strengthen system robustness. Additionally, a fully decoupled detection head substitutes the original coupled head in the YOLOv5s model, separating the object classification and bounding box regression tasks to improve detection accuracy. Finally, a polarization technology-based visual monitoring system is developed to acquire defect images of laser AM workpieces, establishing the model’s training sample database. Compared with YOLOv5, the proposed model demonstrates a 2.5% improvement in precision (P), 2.2% enhancement in recall (R), 3.1% increase in mean average precision (mAP50), and 3.2% elevation in mAP50-95. These quantitative improvements confirm the model’s capability to provide robust and real-time technical solutions for industrial AM quality monitoring, effectively addressing current limitations in defect detection accuracy and system reliability.