Context-Enhanced Network with Spatial-Aware Graph for Smartphone Screen Defect Detection.

Area of Science:

• Computer Vision
• Materials Science
• Deep Learning

Background:

• Touch screen devices are ubiquitous, increasing the need for high-quality screen glass.
• Defect detection in screen glass is critical for smartphone manufacturing quality.
• Existing deep learning methods face challenges with small, irregular, low-contrast defects.

Purpose of the Study:

• To propose an advanced deep learning model for accurate smartphone screen defect detection.
• To address limitations in detecting defects with small size, irregular shapes, and low contrast.

Main Methods:

• Developed CE-SGNet, incorporating an Adaptive Receptive Field Attention Module (ARFAM) and a Spatial-aware Graph Reasoning Module (SGRM).
• ARFAM adaptively extracts contextual information to enhance defect features.
• SGRM uses graph attention networks to encode spatial relationships between defect regions.

Main Results:

• CE-SGNet demonstrated outstanding performance in identifying and locating diverse screen glass defects.
• The model accurately detected defects of various shapes and scales on public datasets.
• Enhanced feature representation improved detection accuracy and robustness.

Conclusions:

• CE-SGNet effectively overcomes challenges in detecting subtle screen glass defects.
• The proposed network significantly advances the state-of-the-art in smartphone screen quality inspection.
• This approach holds promise for improving automated visual inspection systems.