Butterfly and Moth Image Classification using CNN from scratch

This project aims to classify images of butterflies and moths into 10 different species using a custom-built Convolutional Neural Network (CNN) trained from scratch using PyTorch. The project also features an interactive UI using Gradio where users can upload an image and get real-time predictions.

🧠 Pretrained Model Included

This repository includes the saved model weights:
butterfly_model.pth You can reuse the trained model without retraining by loading it as follows:

model = SimpleCNN(num_classes=10) model.load_state_dict(torch.load("butterfly_model.pth", map_location="cpu")) model.eval()

Tools & Libraries Used

• Python, PyTorch
• Torchvision
• Gradio (UI)
• Matplotlib & Sklearn (metrics)
• Google Colab (CPU environment)

Dataset

The dataset was manually curated from a 100-class butterfly/moth dataset. We selected 10 unique species and structured the dataset as follows: my_10_class_dataset/ ├── TRAIN_BALANCED/ # Balanced training set ├── VALID/ # Validation set (5 images per class) └── TEST/ # Test set (5 images per class)

• Image size: 224 x 224 x 3 (RGB)
• Total classes: 10
• Train samples: 1210
• Test/Validation samples: 50 each

Sample Classes

1. BANDED TIGER MOTH
2. CHALK HILL BLUE
3. EMPEROR GUM MOTH
4. GLITTERING SAPPHIRE
5. GREEN HAIRSTREAK
6. MOURNING CLOAK
7. POPINJAY
8. PURPLISH COPPER
9. RED CRACKER
10. ROSY MAPLE MOTH

Class Imbalance Handling

The original training dataset had imbalanced classes. To ensure fairness, we:

• Calculated the minimum class count
• Performed random undersampling of all classes to match the smallest one
• Stored the balanced data in a new folder: TRAIN_BALANCED/

Data Preprocessing & Augmentation

Applied using torchvision.transforms:

• Resize(224, 224) to normalize image size
• RandomHorizontalFlip() & RandomRotation(10) for augmentation
• ToTensor() to convert image to PyTorch tensor
• Normalize([0.5], [0.5]) for input scaling

Model Architecture

Built from scratch using PyTorch:

• 3 Convolutional blocks with ReLU, BatchNorm, and MaxPooling
• Flatten → Dropout → Linear (128) → ReLU → Linear (10 classes)
• Regularization using Dropout (0.5) and BatchNorm

Loss & Optimization:

• Loss: CrossEntropyLoss
• Optimizer: Adam
• Batch size: 16, Epochs: 10, Device: CPU

Model Evaluation

• Evaluation metrics on the test set:
  • Accuracy: 92%
  • Precision, Recall, F1-score per class using classification_report()
  • Confusion matrix (visualized using matplotlib)

Model Saving & Inference

After training:

• Saved model using torch.save(model.state_dict(), 'butterfly_model.pth')
• Reloaded model for inference in Colab

Interactive UI with Gradio

Built a lightweight inference UI using gr.Interface():

• Users can upload any butterfly/moth image
• The trained model returns the predicted class
• Runs directly inside Google Colab

How to Use:

1. Upload your image using the Gradio UI
2. The model will return the butterfly/moth class name
3. Behind the scenes:
   • Image is preprocessed
   • Forward pass through CNN
   • Predicted label returned

Future Improvements

• Use pre-trained models like ResNet for higher accuracy
• Handle class imbalance with SMOTE or weighted loss
• Deploy on HuggingFace Spaces or a Flask web server