Chip Pad Inspection Method Based on an Improved YOLOv5 Algorithm.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Semiconductor Manufacturing

Background:

• Chip pad inspection is crucial for automated semiconductor manufacturing.
• Accurate detection of small chip pads with lightweight models is challenging.
• Existing attention mechanisms struggle with local feature extraction for small targets in complex backgrounds.

Purpose of the Study:

• To develop an effective attention module for small target detection in chip pad inspection.
• To improve the accuracy and efficiency of deep learning-based chip pad inspection models.
• To propose a design guideline for attention layers to optimize network performance and resource usage.

Main Methods:

• Proposed an Object Convolution Attention Module (OCAM) to establish long-range dependencies between channel and position features.
• Integrated the OCAM into the feature extraction layer of the YOLOv5 network.
• Developed a network scaling guideline for attention layers to balance performance and resource requirements.

Main Results:

• The OCAM significantly improved the detection performance of chip pads within the YOLOv5 framework.
• The proposed method demonstrated superior generalization and performance across chip pad, VOC, and COCO datasets.
• Network scaling guidelines helped optimize parameters and reduce hardware demands.

Conclusions:

• The OCAM attention module enhances YOLOv5 for precise and efficient chip pad inspection.
• The approach offers a robust solution for small target detection in complex industrial settings.
• The study provides a valuable methodology for optimizing deep learning models in semiconductor manufacturing.