Fruit Classification Using MobileNetV2

Project Overview

This project aims to classify 10 different types of fruits from images. By leveraging a deep learning model with transfer learning from a pre-trained convolutional neural network, MobileNetV2, the goal is to build an accurate image classification system that can distinguish between various fruit categories. This has applications in automated sorting, inventory management, and computer vision systems for retail.

---

Technical Highlights

• Dataset: Kaggle - Fruit Classification (10 class)
• Size: The dataset contains 10 classes of fruits. It is split into a training set of 1841 images, a validation set of 460 images, and a test set of 1025 images.
• Key Features: The raw fruit image data is used as input for the model.
• Approach:
  • Data Preparation: The images were loaded and organized into training, validation, and test sets using ImageDataGenerator. This approach automatically handles batching and resizing.
  • Data Augmentation: ImageDataGenerator was used to augment the training data with transformations like rotation, shifting, and horizontal flips to increase the dataset's variability and prevent overfitting.
  • Transfer Learning: A pre-trained MobileNetV2 model, which was trained on a large image dataset (ImageNet), is used as a feature extractor. The base model's layers were frozen (trainable=False).
  • Model Architecture: A custom deep learning model was built on top of the MobileNetV2 base, featuring BatchNormalization, Dropout layers, and Dense layers for classification. The final Dense layer has 10 units with a softmax activation for multi-class classification.
  • Training: The model was compiled with the Adam optimizer and categorical_crossentropy loss. It was trained for 80 epochs.
• Best Accuracy:
  • The model achieved a training accuracy of ~98.1% and a validation accuracy of 93.9% in the best epoch. The final evaluation on the validation set resulted in a loss of ~0.26 and an accuracy of 91.8%.

---

Purpose and Applications

• Automated Fruit Sorting: Enables the automatic sorting of fruits in agricultural and food processing facilities.
• Quality Control: Assists in grading produce based on visual characteristics.
• Retail Management: Supports automated checkout systems and inventory management in grocery stores.
• Computer Vision: Provides a practical example of transfer learning for image classification tasks with a moderately sized dataset.

---

Installation

Clone the repository and extract the data from the zip file.

Install the necessary libraries:

pip install tensorflow keras pandas numpy seaborn matplotlib scikit-learn

---

Collaboration

We welcome contributions to improve the project. You can help by:

• Fine-tuning the hyperparameters of the custom head and training process to further optimize performance.
• Exploring different pre-trained models to compare their effectiveness.
• Implementing more robust data augmentation or preprocessing techniques.
• Adding a more detailed evaluation on the test set, including a classification report and confusion matrix, to validate the model's performance on unseen data.