Semiconductor cosmetic defect detection through deep learning

Abstract

Abstract: Defect detection in semiconductor packaging remains a critical challenge due to the high visual complexity of metallic surfaces and the scarcity of defective samples in controlled manufacturing environments. Traditional Automated Optical Inspection (AOI) systems often lack the flexibility to identify subtle cosmetic blemishes, such as micro-scratches and spots, which directly impact product reliability and customer perception. This thesis proposes a robust automated defect detection framework that integrates state-of-the-art Deep Learning (DL) architectures with a distributed Industrial Internet of Things (IIoT) infrastructure.The methodology involves a comprehensive evaluation of three distinct object detection paradigms: two-stage (Faster R-CNN), anchor-free (FCOS), and the latest onestage iteration, YOLOv11. To address the inherent class imbalance, a synthetic data augmentation pipeline was implemented, following recent advancements in generative and instance-level augmentation. Experimental results demonstrate that YOLOv11 achieves a superior balance between accuracy and speed, reaching a mean Average Precision (mAP@0.5) of 93.5% and a precision of 90.7% on an independent test set. Furthermore, the system was validated within an IIoT ecosystem using the MQTT protocol and a Raspberry Pi 5 gateway. These findings suggest that the proposed framework can be seamlessly integrated into high-speed production lines, providing a scalable and high-precision alternative to manual visual inspection in semiconductor back-end manufacturing.