Smartphone screen surface defect detection using dynamic large separable kernel attention and multi-scale feature...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Manufacturing Technology

Background:

• Defect detection is critical for smartphone cover glass quality.
• Complex production environments with reflections pose challenges for current methods.

Purpose of the Study:

• To develop an enhanced YOLOv8 model, DY-YOLO, for accurate and efficient smartphone cover glass defect detection.
• To improve detection accuracy and reduce false positives in challenging industrial settings.

Main Methods:

• Proposed DY-YOLO model integrating Dynamic-Large Separable Kernel Attention (Dynamic-LSKA) for background interference suppression.
• Incorporated Dynamic-C2f module for enhanced feature extraction and Advanced Screening Feature Bidirectional Path Aggregation Network (HSF-BPAN) for feature fusion.
• Utilized DySample as a lightweight dynamic up-sampler to optimize computational cost.

Main Results:

• DY-YOLO achieved state-of-the-art detection accuracy on MSD and SSGD benchmarks, surpassing baseline and existing methods.
• Demonstrated significant improvements in mean Average Precision (mAP) on both datasets.
• Achieved high inference speed (121.8 FPS) with reduced computational cost (33.3% lower) and comparable parameter count to the baseline.

Conclusions:

• DY-YOLO effectively detects cover glass defects in complex industrial environments, offering high accuracy and efficiency.
• The model's performance and speed indicate strong potential for real-time edge deployment in manufacturing quality control.
• The proposed attention module and network architecture contribute to overcoming challenges like glass reflections and background noise.