Snowy-VGG

Overview

Features

• Pretrained Architecture: Built on VGG-19, pretrained on ImageNet.
• Transfer Learning: Fine-tuned for snow classification using a custom dataset.
• Ensemble Capability: Combines with ResNet-50 in the research to improve accuracy and F1 scores.
• Custom Dataset: Trained and validated on a dataset of 98 images of pavements captured via smartphone.
• Optimized for Accuracy: Designed to handle challenging classification scenarios with limited data.

Dataset

• 98 Images: Captured in Minnesota, USA, using a Google Pixel 6a smartphone.
• Resolution: 3024 x 3024 pixels, resized to 128 x 128 for training.
• Labels: Balanced categories of snow and no-snow images.
• Test Set: Includes 22 unseen images from distinct locations to evaluate generalization.

Implementation

• Framework: PyTorch
• Training:
  • Optimizer: Adam
  • Learning Rate: 0.0001
  • Batch Size: 4
  • Epochs: 15-25 (best results at epoch 15)
  • Normalization: Based on ImageNet mean ([0.485, 0.456, 0.406]) and standard deviation ([0.229, 0.224, 0.225]).
• Evaluation Metrics:
  • Accuracy: 72.7% (ensemble)
  • F1 Score: 71.8% (ensemble)

Usage

1. Installation:
   • Install PyTorch and dependencies: pip install torch torchvision
2. Load the Model:
   python

   1import torch 2from torchvision import models 3 4model = models.vgg19(pretrained=True) 5# Modify the classification layer for binary output (snow vs no-snow) 6model.classifier[-1] = torch.nn.Linear(model.classifier[-1].in_features, 2) 7model.load_state_dict(torch.load('Best_Model_VGG19.pt')) 8model.eval()
3. Inference:
   • Preprocess the input image to 128x128 pixels.
   • Normalize using ImageNet statistics.
   • Pass the image through the model for predictions:
     python

     1with torch.no_grad(): 2 output = model(image_tensor) 3 prediction = torch.argmax(output, dim=1) 4 print("Snow detected" if prediction.item() == 1 else "No snow detected")

Results

• Performance:
  • Ensemble of Snowy-VGG and ResNet-50 achieves an F1 Score of 71.8% and an accuracy of 72.7% on unseen test data.
  • Focuses on reducing false negatives to enhance pedestrian safety.

Limitations

• Limited Dataset: Only 98 images for training, which may affect generalization.
• Light Reflection Sensitivity: Difficulty in detecting subtle changes in light wave reflections on snow.
• False Negatives: High-risk misclassification when snow is not detected.

Future Directions

• Dataset Expansion: Include more diverse images from different locations, lighting conditions, and camera types.
• Model Enhancements: Incorporate custom layers to better handle light reflection variations.
• Real-Time Applications: Adapt Snowy-VGG for mobile deployment to provide real-time alerts for pedestrians.

Citation

Ricardo de Deijn and Rajeev Bukralia, "Image Classification for Snow Detection to Improve Pedestrian Safety," Midwest Association for Information Systems (MWAIS) 2024. Link to Paper

Acknowledgments

• Ricardo de Deijn: ricardodedeijn@gmail.com