Fruit Detection

Overview

This project builds a CNN to classify 255 classes of fruits, vegetables, nuts, and seeds from the Fruits-360 dataset. It includes analysis such as a confusion matrix, t-SNE visualization, and Grad-CAM to understand how the model makes predictions.

Setup

Option 1: Google Colab

1. Open main_submission.ipynb in Google Colab
2. Set RUN_ON_COLAB = True
3. Enable GPU (Runtime > Change runtime type > GPU)
4. Run all cells

Note: Google Drive access is required to save outputs.

Option 2: Local Environment

pip install -r requirements.txt # install dependencies
jupyter notebook main_submission.ipynb 

Dataset

• Source: Fruits-360 Dataset
• Classes: 255 (fruits, vegetables, nuts, seeds)
• Download method: Automatic via kagglehub

Model

• Architecture: CNN with Conv, BatchNorm, ReLU, and Pooling layers
• Input size: 64 × 64
• Optimizer: Adam (learning_rate = 0.001)
• Loss function: SparseCategoricalCrossentropy

Results

Metric	Value
Test accuracy	96.66%
Precision	96.85%
Recall	96.66%
F1-score	96.59%

The model shows lower performance on similar classes (e.g., different blackberry variants).

Analysis

• Confusion Matrix: Errors occur between similar classes.
• t-SNE Visualization: Classes form distinct clusters in feature space.
• Grad-CAM: Model attention aligns with fruit regions in images.

Repo Structure

fruit_detection/
├── main_submission.ipynb
├── requirements.txt
├── README.md
└── fruit_model/

Notes

• GPU is recommended for training
• Image size is 64×64 due to memory constraints
• steps_per_epoch = 2000 (partial data per epoch)
• Dataset caching is not used due to memory constraints

License

MIT License