plant-detector

Model Description

Convolutional Autoencoder for plant anomaly detection

This is a Convolutional Autoencoder (CAE) trained for plant anomaly detection. The model learns to reconstruct plant images and detects anomalies based on reconstruction error.

Model Details

• Model Type: Convolutional Autoencoder
• Framework: PyTorch Lightning
• Task: Anomaly Detection / Plant Classification
• Input: RGB images (224x224)
• Output: Reconstruction + anomaly score

Training Details

• Architecture: Encoder-Decoder with skip connections
• Loss Function: Mean Squared Error (MSE)
• Optimizer: AdamW
• Learning Rate: 0.0001
• Batch Size: 32
• Epochs: N/A
• Dataset Size: N/A images

Performance Metrics

• Validation Loss: N/A
• Threshold: 0.5687
• Mean Reconstruction Error: N/A
• Std Reconstruction Error: N/A
• Anomaly Rate: N/A

Normalization Statistics

The model expects input images to be normalized with:

• Mean: [0.4682, 0.4865, 0.3050]
• Std: [0.2064, 0.1995, 0.1961]

Usage

PyTorch Lightning Checkpoint

from annomallyDet.models.lit_models.lit_cae import LitCAE

# Load the model
model = LitCAE.load_from_checkpoint("plant_anomaly_detector.ckpt")
model.eval()

# Make prediction
reconstruction_error = model.get_reconstruction_error(input_tensor)
is_anomaly = reconstruction_error > 0.5687

Mobile Deployment (TorchScript Lite)

import torch

# Load mobile model
model = torch.jit.load("plant_anomaly_detector.ptl")
reconstruction = model(input_tensor)

# Calculate reconstruction error
error = torch.mean((input_tensor - reconstruction) ** 2)
is_anomaly = error > 0.5687

Flutter Integration

See the included flutter_integration_example.dart for complete Flutter app integration using flutter_pytorch_lite.

Files Included

• plant_anomaly_detector.ckpt: PyTorch Lightning checkpoint
• plant_anomaly_detector.ptl: TorchScript Lite model for mobile deployment
• config.json: Model configuration and metadata
• flutter_integration_example.dart: Flutter integration example
• normalization_stats.json: Dataset normalization statistics

Model Architecture

Input (3, 224, 224)
    ↓
Encoder: Conv2d → BatchNorm → LeakyReLU → Dropout
    [32, 64, 128, 256] channels
    ↓
Latent Space (128 dimensions)
    ↓
Decoder: ConvTranspose2d → BatchNorm → LeakyReLU → Dropout
    [256, 128, 64, 32] channels
    ↓
Output (3, 224, 224)

Anomaly Detection Logic

1. Training: Model learns to reconstruct normal plant images
2. Inference: Calculate reconstruction error (MSE)
3. Decision: If error > threshold → Anomaly (not a plant)
4. Confidence: Distance from threshold indicates confidence

Limitations

• Trained specifically on plant images
• Performance depends on similarity to training data
• May struggle with novel plant species not in training set
• Threshold may need adjustment for different use cases

Citation

@misc{plant_anomaly_detector,
  title={Plant Anomaly Detection using Convolutional Autoencoder},
  author={Your Name},
  year={2024},
  howpublished={\url{https://huggingface.co/YOUR_USERNAME/plant-detector}},
}

License

[Specify your license here]