Model Card for Model ID

A custom cnn base model trained specifically to classify images of fruits and vegetables.

Model Details

Model Description

Model Name: Fruit-Veg-Img-Classifier
Model Type: Image Classification
Framework: TensorFlow
Language: Python, Jupyter Notebook
License: MIT
Author: Senadhi Chandrasekara
Repository: GitHub
Demo: Streamlit App

Overview

The Fruit-Veg-Img-Classifier is a deep learning-based Convolutional Neural Network (CNN) designed to classify images into 36 categories of fruits and vegetables. This model was initially inspired by a tutorial by KothaEd on YouTube and has been enhanced with custom modifications for improved performance and usability. It is trained on a dataset of approximately 3,825 images, achieving a Top-3 Accuracy of 0.96657383.

Model Architecture

Input: Images resized to 180x180 pixels.
Layers:
- Rescaling layer (normalizes pixel values to [0, 1]).
- Three Conv2D + MaxPooling2D blocks (16, 32, 64 filters).
- Flatten layer.
- Dropout (0.2) for regularization.
- Dense layer (128 units) followed by an output layer (36 units with softmax activation).
Optimizer: Adam
Loss Function: Sparse Categorical Crossentropy
Training: 39 epochs with data augmentation (random flips, rotations, zooms).
Metrics: Accuracy, Top-3 Accuracy

Training Data

Dataset: Custom dataset of 36 fruit and vegetable classes (e.g., apple, banana, beetroot, etc.).
Split:
- Train: ~3,115 images
- Validation: ~351 images
- Test: ~359 images
Source: Google Drive Dataset
Preprocessing: Resizing, shuffling, batching, and data augmentation.

Performance

Training Accuracy: ~95-98%
Validation Accuracy: ~85-95%
Top-3 Accuracy: 0.96657383
Evaluation Metrics: Includes confusion matrix, precision, recall, and F1 score (computed via model-evaluation.ipynb).

Usage

This model is deployed as a web application using Streamlit, accessible online. Users can input images via URL pasting, drag-and-drop, or file browsing. The model predicts the fruit or vegetable class with a confidence score (e.g., "Fruit in image is apple with an accuracy of 99.95%").

Installation

pip install tensorflow streamlit numpy pandas matplotlib huggingface_hub

How to Use

Load the model from Hugging Face using an access token stored in a .env file.
Integrate with the provided app.py for web deployment or use image-classification-model-refine.ipynb for local training.
Example inference can be run locally after cloning the repository: git clone https://github.com/senadhi-Thimanya/image-classification-neural-network-model.git.

Use the code below to get started with the model.

import numpy as np
import tensorflow as tf
from tensorflow import keras

model = tf.keras.models.load_model('Image_classify.keras')
data_train_path = 'Fruits_Vegetables/train'
data_train = tf.keras.utils.image_dataset_from_directory(
    data_train_path,
    shuffle=True,
    image_size=(180,180),
    batch_size=32,
    validation_split=False)
data_cat = data_train.class_names #download the dataset

image = 'corn.jpg'
image = tf.keras.utils.load_img(image, target_size=(180,180))
img_arr = tf.keras.utils.array_to_img(image)
img_bat=tf.expand_dims(img_arr,0)

predict = model.predict(img_bat)

score = tf.nn.softmax(predict)

print('Veg/Fruit in image is {} with accuracy of {:0.2f}'.format(data_cat[np.argmax(score)],np.max(score)*100))

Limitations

Requires internet access for Hugging Face model loading in the web app.
Performance may vary with images significantly different from the training dataset.
No support for local file I/O in the web app due to Streamlit constraints.

Credits

Original Tutorial: KothaEd (YouTube: Image Classification Deep Learning Neural Network Model in Python with TensorFlow).
Enhancements: Custom training (39 epochs), data augmentation, Hugging Face integration, and multi-input web app features by senadhi thimanya.

Citation

If you use this model, please cite the GitHub repository and acknowledge the original tutorial by KothaEd.

Model Card Authors

Senadhi Chandrasekara

Model Card Contact

[email protected]

Downloads last month: 855