Model Card for Industrial Tile Defect Detection (YOLO11-Large)

This model is a fine-tuned version of YOLO11-Large designed to detect minute manufacturing defects on industrial tiles. It was trained using a sliding-window tiling approach with aggressive oversampling to handle severe class imbalance.

Model Details

Model Description

This model solves the problem of detecting tiny defects (approx. 10-50 pixels) on massive industrial images (8192x6000 pixels). Standard resizing destroys these defects, so this model was trained on 640x640 crops generated from the original high-resolution data.

To address the rarity of specific defects (like "Halos"), the training pipeline utilized a dynamic oversampling strategy where rare defect crops were physically duplicated with random jitter, ensuring the model saw a balanced distribution of classes.

• Developed by: Can Deniz Kocak
• Model type: Object Detection (YOLO11)
• Finetuned from model: yolo11l.pt (YOLO11 Large)
• License: Apache 2.0

Model Sources

• Repository: huggingface.co/candenizkocak/tile-defect-detection-yolo11
• Framework: Ultralytics YOLO

Uses

Direct Use

The model is intended for Industrial Quality Control (QC) systems. It accepts 640x640 image tensors.

Important: Because the original images are high-resolution (4K/8K), this model should be used within a Sliding Window Inference pipeline. Running it directly on a downscaled 8000px image will result in poor performance.

Classes Detected

The model detects 6 specific defect categories:

1. Edge defect (ID: 0)
2. Corner defect (ID: 1)
3. White spot (ID: 2)
4. Light patch (ID: 3)
5. Dark spot/patch (ID: 4)
6. Halo (ID: 5)

Out-of-Scope Use

• This model is not suitable for general object detection (people, cars, etc.).
• It is optimized for gray/textured industrial tile backgrounds. Using it on different materials (wood, metal) may require fine-tuning.

How to Get Started with the Model

You can use this model directly with the ultralytics library.

from ultralytics import YOLO
import cv2

# Load the model
model = YOLO("https://huggingface.co/candenizkocak/tile-defect-detection-yolo11/resolve/main/best.pt")

# Inference on a single image (or crop)
# Note: For large images, use a sliding window approach
results = model.predict("path/to/tile_crop.jpg", conf=0.35)

# Visualize
results[0].show()

Training Details

Training Data

The dataset consists of 5,388 high-resolution industrial images.

• Preprocessing: Images were sliced into 640x640 tiles.
• Augmentation: Random crop jittering was applied to duplicate defects.

Training Procedure

To combat the massive class imbalance (e.g., ~9000 Dark Spots vs ~300 Halos), an Aggressive Oversampling strategy was applied during the tiling phase:

This ensured that the model saw approximately ~8,000 examples of each class during training.

Training Hyperparameters

• Epochs: 50
• Batch Size: 32
• Image Size: 640
• Optimizer: Auto (AdamW)
• Hardware: NVIDIA A100 (80GB)

Evaluation

Metrics

The model achieves strong performance across all classes, with particularly notable success in detecting the rare "Halo" class due to the oversampling strategy.

• mAP@50 (Mean Average Precision): ~78%
• Recall (Dark Spots): ~88%
• Recall (Halos): ~74% (Significantly improved from baseline)

Recommendations

False Positives on Background: The model may occasionally detect defects on the black background (conveyor belt). It is highly recommended to use a Region of Interest (ROI) filter (such as convex hull detection on the tile) to ignore predictions outside the tile area.

Environmental Impact

• Hardware Type: NVIDIA A100-SXM4-80GB
• Hours used: ~2 hours
• Cloud Provider: Google Colab Pro+
• Compute Region: US-Central1

Technical Specifications

Model Architecture

YOLO11 employs a CSP-based backbone with an improved detect head compared to YOLOv8, offering a better trade-off between latency and accuracy for small object detection.

Software

• Ultralytics: 8.3.x
• PyTorch: 2.9.x
• CUDA: 12.x