Image Classification - ResNet

This case study demonstrates training a ResNet-18 convolutional neural network to classify bean plant diseases. ResNet (Residual Network) uses skip connections to enable training of very deep networks, making it highly effective for computer vision tasks.

Dataset: Beans Plant Disease

• Source: HuggingFace (AI-Lab-Makerere/beans)
• Type: Image classification
• Size: 1,295 images
• Classes: 3 (Angular Leaf Spot, Bean Rust, Healthy)
• Resolution: 500x500 pixels
• Format: RGB color images

Model Configuration

{
  "model": "resnet18",
  "category": "computer_vision",
  "subcategory": "image-classification",
  "model_config": {
    "pretrained": true,
    "num_classes": 3,
    "batch_size": 32,
    "epochs": 15,
    "learning_rate": 0.001,
    "optimizer": "adam",
    "image_size": [224, 224]
  }
}

Training Results

Training Progress

Loss and accuracy curves over 15 epochs:

Class Performance

Per-class precision, recall, and F1 scores:

Confusion Matrix

Model predictions vs actual classes:

Sample Predictions

Example predictions with confidence scores:

Common Use Cases

• Agricultural Disease Detection: Early identification of crop diseases
• Medical Imaging: Classify X-rays, MRIs, CT scans
• Quality Control: Detect defects in manufactured products
• Wildlife Monitoring: Identify animal species from camera traps
• Satellite Imagery: Land use classification, deforestation detection
• Retail: Product categorization from images

Key Settings

Essential Parameters

• pretrained: Use ImageNet pre-trained weights (recommended)
• num_classes: Number of output categories
• batch_size: Images per training batch (16-64 typical)
• learning_rate: Step size for optimization (0.0001-0.001)
• epochs: Number of complete passes through dataset

Data Augmentation

• rotation: Random rotation (-20° to +20°)
• flip: Horizontal/vertical flipping
• zoom: Random zoom (0.8x to 1.2x)
• brightness: Brightness adjustment
• normalization: ImageNet statistics (mean, std)

Advanced Configuration

• optimizer: Adam, SGD, or AdamW
• scheduler: Learning rate scheduling (StepLR, CosineAnnealing)
• dropout: Regularization to prevent overfitting
• weight_decay: L2 regularization strength

Performance Metrics

• Overall Accuracy: 97.0% on test set
• Macro Average Precision: 96.3%
• Macro Average Recall: 96.7%
• Macro Average F1: 96.3%
• Training Time: ~18 minutes (GPU: NVIDIA RTX 3080)
• Inference Speed: ~45 images/second

Tips for Success

1. Transfer Learning: Always start with pretrained weights when possible
2. Data Augmentation: Essential for small datasets to improve generalization
3. Proper Splits: Use stratified splits to maintain class balance
4. Learning Rate: Start with 0.001, reduce if training unstable
5. Early Stopping: Monitor validation loss to prevent overfitting
6. Image Quality: Ensure consistent image quality and resolution

Example Scenarios

Scenario 1: Healthy Bean Plant

• Input: Clear image of healthy bean leaves
• Prediction: Healthy (confidence: 99.2%)
• Action: No treatment needed

Scenario 2: Angular Leaf Spot Detected

• Input: Image showing characteristic angular lesions
• Prediction: Angular Leaf Spot (confidence: 96.8%)
• Action: Apply copper-based fungicide, remove affected leaves

Scenario 3: Bean Rust Identification

• Input: Image with rust-colored pustules on leaves
• Prediction: Bean Rust (confidence: 95.4%)
• Action: Apply sulfur-based fungicide, improve air circulation

Troubleshooting

Problem: Model overfitting (high training acc, low validation acc)

• Solution: Increase data augmentation, add dropout, reduce model complexity

Problem: Low accuracy on all classes

• Solution: Train longer, increase model capacity, improve data quality

Problem: One class performing poorly

• Solution: Check class balance, add more samples, adjust class weights

Problem: Predictions inconsistent

• Solution: Ensure proper image preprocessing, check normalization

Model Architecture Highlights

ResNet-18 consists of:

• Initial Conv Layer: 7x7 convolution, 64 filters
• 4 Residual Blocks: Progressive feature extraction
• Skip Connections: Enable gradient flow in deep networks
• Global Average Pooling: Reduce spatial dimensions
• Fully Connected Layer: Final classification (3 classes)
• Total Parameters: ~11.7 million (11.2M pretrained + custom head)

Next Steps

After training your ResNet-18 model, you can:

• Deploy as REST API for real-time predictions
• Create mobile app using TensorFlow Lite or PyTorch Mobile
• Integrate with drone imaging for large-scale crop monitoring
• Fine-tune on additional disease types
• Export to ONNX for cross-platform deployment
• Compare with larger models (ResNet-50, EfficientNet)