At Ansell Lanka, I contributed to the development and testing of a real-time defect detection system for medical glove packaging lines using YOLO-based object detection. The system — known internally as the Outer Pack Automated Optical Inspection (AOI) Project — was integrated with MultiVac packing machines across three supermarket packaging lines: SPL2, SPL3, and SPL4.
System Architecture
The deployed system consisted of three tightly integrated components:
- Camera & Electronic Panel — captures real-time images of glove packages on the conveyor belt
- Python-based inference software — classifies defects per group of 12 packages
- Industrial Printer — marks defective packs via Ethernet connection with the IPC (Industrial Personal Computer), enabling targeted removal
Defect Classes
The system was designed to classify three types of defects in medical glove outer packaging:
- Top print smudging — unclear or smudged print on the top surface of the package
- Bottom print smudging — unclear or smudged print on the bottom surface
- Seal integrity failures — including air bubbles inside the seal, packaging material caught in the seal, and text printed incorrectly on the seal region
Key Technical Contributions
Data Preprocessing & Lightning Equalization
Implemented image preprocessing to equalize lighting conditions across camera frames. This was critical to prevent the thresholding algorithm (used for caught-in-seal detection) from producing false positives due to lighting variation across the packaging line.
Synthetic Dataset Generation
Air bubble defects inside seals were rare in production, creating a significant class imbalance problem. To address this:
- Extracted defect regions from the limited real images using graphics software
- Wrote a Python script to composite and blend extracted defects onto clean seal images at varied positions and opacities
- Generated a balanced synthetic dataset sufficient to train a robust YOLOv8 model
YOLOv8 Model Training — Seal Integrity Detection
Trained and validated a YOLOv8 object detection model specifically for detecting air bubble defects in seal regions. The model was trained on the synthetically augmented dataset and evaluated against real production samples.
Text Detection & Erasure Pipeline
Used a pretrained EasyOCR model to detect text regions on seal areas. Detected text bounding boxes were erased from the image before the caught-in-seal thresholding step, preventing false detections caused by printed text patterns. This pipeline was later optimized for speed to meet real-time performance requirements.
Rule-Based Empty Pack Detection
Implemented a rule-based algorithm to distinguish between empty packages and those containing gloves — a prerequisite for accurate defect classification downstream.
Socket Programming — Industrial Printer Interface
Developed a Python class to manage Ethernet communication between the IPC and the industrial printer. The class exposed functions for:
- Initiating and maintaining connections
- Reading from and writing to the printer
- Updating the print job dynamically for each group of 12 packs based on software inference results
Synchronization Testing
Conducted two-phase synchronization testing to verify that defect marks were printed on the exact package flagged by the model:
- Replaced model inferences with known test patterns and validated printer output
- End-to-end testing with live model inferences on the packaging line
Performance Optimization
Profiled and optimized the multithreaded software architecture to eliminate bottlenecks. Drew time diagrams to trace processing time across pipeline stages. Key process times measured after optimization (seconds):
| Pipeline Stage | Before | After |
|---|---|---|
| Defect on top & return coords | 2.0865 | 0.66415 |
| Defect on bottom & return coords | 1.0252 | 0.1191 |
| Correct position | 0.011 | 0.0346 |
| Total | — | ~0.823 |
Physical Validation
Conducted side-by-side comparison between the system’s defect classifications and those made by experienced production line workers to validate real-world accuracy and calibrate rejection thresholds.
Technologies & Tools
Libraries & Frameworks:
- Python, PyQt (GUI)
- YOLOv8 (object detection)
- EasyOCR (pretrained OCR model)
- OpenCV, Pillow (image processing)
- NumPy, Keras (ML utilities)
- CUDA (GPU acceleration)
Development Tools:
- PyCharm IDE
- Visual Studio IDE
Impact
- Automated manual defect inspection across three packaging lines, reducing reliance on human visual inspection
- Achieved real-time inference within ~0.82 seconds per batch of 12 packages after optimization
- Contributed to a research paper published by the Ansell AI Research Team on this system
- Authored technical documentation including a user manual and IQ (Installation Qualification) document
- Assisted in training production staff and replicating the development environment for the SPL3 line
