Real-Time 3D Reconstruction for Wire Bonding Using Multi-View Projection and EM Polynomial Modelling

Automated Optical Inspection (AOI) systems are critical in ensuring wire bonding quality during integrated circuit (IC) packaging. As wire diameters shrink and interconnect layouts become more complex, conventional 2D inspection faces challenges such as occlusion and overlap, leading to missed defect detections. Although 3D inspection methods can improve accuracy, they are often costly and reduce throughput in inline metrology. This paper proposes a lightweight, projection-based framework that combines parametric polynomial modelling with an Expectation-Maximization (EM) optimization strategy to achieve accurate and real-time 3D wire reconstruction from standard multi-view grayscale images. The method models wire trajectories as piecewise cubic polynomials, jointly fitted across multiple views using known projection matrices. An iterative EM procedure refines wire parameters under partial occlusion, image noise, and missing data. Experiments on industrial AOI datasets demonstrate that the proposed method reconstructs each wire within <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{3 0 ~ m s}$</tex>, achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim \text{1 5. 5 5 ~} \mu$</tex> m loop height accuracy, and maintains over <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$90 \% 2 \mathrm{D}$</tex> coverage, making it a practical and scalable solution for high-speed, high-accuracy wire bond inspection.