YOLOSeg with applications to wafer die particle defect segmentation

Area of Science:

• Computer Vision
• Materials Science
• Semiconductor Manufacturing

Background:

• Accurate detection of small particle defects on wafer dies is crucial for semiconductor quality control.
• Existing instance segmentation models often struggle with the scale and subtlety of these defects.
• Efficient training data preparation is a significant bottleneck in developing robust defect detection systems.

Purpose of the Study:

• To develop an end-to-end instance segmentation model, YOLOSeg, for precise identification and segmentation of small particle defects on wafer dies.
• To enhance the model's performance through optimized training strategies and advanced augmentation techniques.
• To demonstrate YOLOSeg's superiority over current state-of-the-art methods in terms of accuracy and efficiency.

Main Methods:

• Development of YOLOSeg, an instance segmentation model based on YOLOv5s, incorporating a UNet-like segmentation head.
• Implementation of training optimizations including layer freezing, mask loss switching, auto-anchor utilization, and denoising diffusion probabilistic models (DDPM) for image augmentation.
• Evaluation of YOLOSeg's performance using average precision (AP) and intersection over union (IoU) metrics on a dedicated testing dataset.

Main Results:

• YOLOSeg achieved high performance metrics, with an average precision (AP) of 0.821 and an intersection over union (IoU) of 0.732.
• The model demonstrated superior performance in segmenting extremely small particle defects compared to established models like Mask R-CNN, YOLACT, YUSEG, and YOLOv5s-segmentation.
• The training process for YOLOSeg requires less time and effort, eliminating the need for extensive data collection, pseudo-defect annotation, or complex feature engineering.

Conclusions:

• YOLOSeg is a highly effective and efficient model for instance segmentation of small particle defects in semiconductor manufacturing.
• The proposed training strategies and DDPM augmentation significantly improve segmentation accuracy, especially for challenging small-scale defects.
• YOLOSeg offers a practical and time-saving solution for wafer defect inspection, reducing the burden of data preparation and enhancing quality control.