MSAdaNet: An Adaptive Multi-Scale Network for Surface Defect Detection of Smartphone Components

Area of Science:

• Computer Vision
• Industrial Manufacturing
• Machine Learning

Background:

• Surface defect detection is crucial for smartphone quality assurance.
• Existing deep learning methods face challenges with diverse defect morphology and limited labeled data.

Purpose of the Study:

• To develop an effective deep learning solution for industrial surface defect detection.
• To address the scarcity of labeled training data in defect detection tasks.

Main Methods:

• Introduction of MSAdaNet (Multi-Scale Adaptive Defect Detection Network) with novel PMSFA backbone, FDPN neck, and SASD head.
• Development of a synthetic data generation pipeline to create the Smartphone Camera Bezel Dataset (SCBD).

Main Results:

• MSAdaNet achieved a state-of-the-art mAP@0.5 of 54.8% on the SSGD dataset, outperforming existing frameworks.
• Achieved 94.0% mAP@0.5 on the synthetic SCBD, demonstrating the effectiveness of the data generation pipeline and model robustness.
• Ablation studies confirmed the significant contribution of each proposed module.

Conclusions:

• MSAdaNet offers an effective and efficient solution for industrial surface defect detection.
• The proposed data generation pipeline successfully addresses data scarcity challenges.
• The developed network architecture demonstrates robustness across different data distributions.