Small object detection method with shallow feature fusion network for chip surface defect detection.

Area of Science:

• Intelligent manufacturing
• Semiconductor industry
• Computer vision

Background:

• Intelligent manufacturing emphasizes flexibility, customization, and quality control, critical for chip production.
• Automated industrial lines face challenges in intelligent chip defect detection and classification.
• Existing object detection methods struggle with small targets, particularly chip surface defects.

Purpose of the Study:

• To enhance the performance of chip defect detection, especially for small objects.
• To address the limitations of current methods in identifying subtle defects on chip surfaces.
• To improve the accuracy and efficiency of defect detection in semiconductor manufacturing.

Main Methods:

• Proposed a small object detection method based on YOLOv4, termed SO-YOLO.
• Expanded feature fusion of shallow features to capture finer details.
• Optimized anchor box dimensions using k-means++ clustering.
• Increased detection efficiency by removing redundant YOLO head network branches.

Main Results:

• SO-YOLO demonstrated superior performance compared to original YOLOv4, YOLOv5s, and YOLOv5l models.
• Achieved improvements in parameter count, classification accuracy, and detection accuracy.
• Effectively enhanced the detection of small defects on chip surfaces.

Conclusions:

• The proposed SO-YOLO method significantly improves small object detection for chip defect analysis.
• SO-YOLO offers a more efficient and accurate solution for quality control in semiconductor manufacturing.
• This advancement contributes to more robust intelligent manufacturing processes in the electronics industry.