AOI System for Apple iPhone Front Glass Surface Inspection

14 defect classes · YOLOv11 + SegNExt · TensorRT optimization · Real-time production deployment

Overview

Advanced AOI system for iPhone front glass surface inspection detecting 14 types of defects including scratches, particles, contamination, pinholes, and surface anomalies. Hybrid approach combining object detection (YOLOv11) and segmentation (SegNExt) optimized with TensorRT for real-time manufacturing deployment.

🎯 Object Detection 🔬 Segmentation ⚡ TensorRT Python PyTorch OpenCV 🅲# C#

Problem

  • 14 different defect types requiring both detection and precise localization.
  • High-throughput production requires real-time inference speed.
  • Small defects (pinholes, particles) need pixel-level accuracy.
  • Reflective glass surfaces create challenging inspection conditions.

Solution

  • Hybrid inspection: YOLOv11 for fast detection + SegNExt for precise segmentation.
  • Multi-class defect recognition covering all 14 defect categories.
  • TensorRT optimization for real-time performance in production.
  • Custom training with augmentation for robustness against reflections.

My Role

  • Designed and integrated hybrid inspection pipeline (detection + segmentation).
  • Customized YOLOv11 and SegNExt architectures for multi-class defect recognition.
  • Applied TensorRT optimization for real-time deployment.
  • Validated system performance under production constraints and variations.

Implementation

Languages: Python & C#

Libraries:

PyTorch ⚡ TensorRT OpenCV

Models: YOLOv11, SegNExt

Defect Classes: Scratches, particles, contamination, pinholes, cracks, chips, bubbles, and more (14 total)

Results Gallery

Training and evaluation demonstrations are omitted due to data confidentiality.

Demo Video

Training and evaluation demonstrations are omitted due to data confidentiality.

End Result

  • Fast and highly accurate multi-class inspection system (14 defect types).
  • Production-grade speed and robustness for high-throughput manufacturing.
  • Real-time TensorRT-optimized inference suitable for inline inspection.
  • Deployed successfully in Apple manufacturing environment.