Fruit Image Classifier

A simple web-app for fruit image classification.

This simple web-app identifies and classifies a digital image of a fruit amongst six fruit categories: Apple, Avocado, Banana, Kiwi, Orange. The app uses a trained CNN model, as its prediction engine, which has a classification accuracy of 92% on the test dataset.

Structure of the repository

|-- Model                             # contains the trained CNN model
|-- app_interface
|-- python - code
|   |-- archive models                # cotains old python code files for training CNN
|   |-- cnn_tuned_architecture.py     # final code used for training the model
|   app.py                            # streamlit app
|-- data.zip                          # fruit classification data set in zip format
|-- requirements.txt                  # list of dependencies and version numbers
|-- utility_functions.py              # Some functions used by the streamlit app

Data

The data is in zip format. Source: The image dataset has been mainly sourced from an online source. When unzipped, the structure of the data folder is as following:

|-- data
|   |-- test
|       |-- apple
|       |-- avocado
|       |-- banana
|       |-- kiwi
|       |-- organge
|   |-- training
|       |-- apple
|       |-- avocado
|       |-- banana
|       |-- kiwi
|       |-- organge
|   |-- validation
|       |-- apple
|       |-- avocado
|       |-- banana
|       |-- kiwi
|       |-- organge

Usage:

(Before using the app, make sure you have all the depencies installed as listed in the requirements.txt)

• Pull the repository to local drive
• In the terminal/command prompt, navigate to the directory containing app.py
• Type in the command:

streamlit run app.py

Model performance

The model has been trained on limited dataset which contains clean 'studio-type' images of the five fruit classes with clear background. The validation and test datasets also correspond to the training dataset. As such, the model is able to achieve 98% accuracy on the test set.

The fruit classifier app, however, is unable to achieve 98% accuracy on any images of the fruit having rich background. In future, I intend to re-train the model on even richer fruit images dataset to increase its classification accuracy of fruit images with diverse backgrounds.

Streamlit User Interface