This repo contains code for conducting image classification on a dataset of fruit images. Two models are fit to the data; a simple sequential model which is akin to multiclass logistic regression, and a large pretrained CNN model (VGG16). The project demonstrates why some tasks are best solved using lower complexity models, as is the case for this project.
The dataset used for the project is the Fruit Classification dataset from Kaggle. This dataset consists of 22495 images of fruits across 33 classes (fruit types). Images are 100x100 resolution and masked to isolate the fruits. The samples in each class appear to be quite heterogeneous in this dataset, so a high accuracy score is expected when classifying the images. The data is pre-split into a train and test set, but since the test set doesn't have labels, the training data was furhter split into a training (80%, 11787 images), validation (20%, 3361 images) and test split (10%, 1706 images) with the split-folders package for Python.
Below are examples from each class of fruit:
Image source
Example of Apple Braeburn test set:
Image classification was performed using two approaches:
- Initiating, training and predicting with a simple sequential model.
- Finetuning and predicting with the
VGG16model.
The simple model is a sequential model with zero hidden layers; there is only an input layer fully connected to a 33 node dense output layer. This is essentially multiclass logisitic regression.
The VGG16 model is 16 layers deep, and has approx 138 million parameters and is trained on the ImageNet dataset. (source).
Both models are handled in the TensorFlow framework.
There are two Python scripts in the src folder, simple_fruit_classifier.py and fruit_classifier.py, which contain code pipelines for performing image classification using the two models. Each script follows these steps:
- Import dependencies
- Load and preprocess data
- Setup data generators
- Setup model
- Fit model to data
- Plot and save learning curves
- Print and save classification report
The fruit_classifier.py script uses the VGG16 model.
The code is tested on Python 3.11.2. Futhermore, if your OS is not UNIX-based, a bash-compatible terminal is required for running shell scripts (such as Git for Windows).
The repo was setup to work with Windows (the WIN_ files), MacOS and Linux (the MACL_ files).
git clone https://github.com/alekswael/fruit_classifier_LR_VGG16
cd fruit_classifier_LR_VGG16NOTE: Depending on your OS, run either WIN_setup.sh or MACL_setup.sh.
The setup script does the following:
- Creates a virtual environment for the project
- Activates the virtual environment
- Installs the correct versions of the packages required
- Deactivates the virtual environment
bash WIN_setup.shNOTE: Depending on your OS, run either WIN_run.sh or MACL_run.sh.
Run the script in a bash terminal.
The script does the following:
- Activates the virtual environment
- Runs either
fruit_classifier.pyorsimple_fruit_classifier.pylocated in thesrcfolder - Deactivates the virtual environment
bash WIN_run.sh
bash WIN_run_simple.shSome model parameters can be set through the argparse module. However, this requires running the Python script seperately OR altering the run*.sh file to include the arguments. The Python script is located in the src folder. Make sure to activate the environment before running the Python script.
simple_fruit_classifier.py [-h] [-bs BATCH_SIZE] [-e EPOCHS]
options:
-h, --help show this help message and exit
-bs BATCH_SIZE, --batch_size BATCH_SIZE
Batch size for training. (default: 32)
-e EPOCHS, --epochs EPOCHS
Number of epochs to train for. (default: 100)
fruit_classifier.py [-h] [-bs BATCH_SIZE] [-e EPOCHS]
options:
-h, --help show this help message and exit
-bs BATCH_SIZE, --batch_size BATCH_SIZE
Batch size for training. (default: 32)
-e EPOCHS, --epochs EPOCHS
Number of epochs to train for. (default: 100)
This repository has the following structure:
│ MACL_run.sh
│ MACL_run_simple.sh
│ MACL_setup.sh
│ README.md
│ requirements.txt
│ WIN_run.sh
│ WIN_run_simple.sh
│ WIN_setup.sh
│
├───extra
│ examples.png
│ homogeneity.png
│
├───fruits_v2
│ ├───test
│ │ ├───Apple Braeburn
│ │ │ Apple Braeburn.jpg
│ │ │ ...
│ │ ├───Apple Granny Smith
│ │ │ Apple Granny Smith.jpg
│ │ │ ...
│ │ ...
│ ├───train
│ │ ├───Apple Braeburn
│ │ │ Apple Braeburn.jpg
│ │ │ ...
│ │ ├───Apple Granny Smith
│ │ │ Apple Granny Smith.jpg
│ │ │ ...
│ │ ...
│ ├───val
│ │ ├───Apple Braeburn
│ │ │ Apple Braeburn.jpg
│ │ │ ...
│ │ ├───Apple Granny Smith
│ │ │ Apple Granny Smith.jpg
│ │ │ ...
│ │ ...
│
├───out
│ classification_report.txt
│ classification.png
│ simple_classification.png
│ simple_classification_report.txt
│
└───src
fruit_classifier.py
simple_fruit_classifier.py
The simple sequential model achieves a near perfect score of ~99% acc after 13 epochs when predicting the test set. Val-accuracy is about 96% after just 7 epochs, so further training yields diminishing returns after this point.
Upon reviewing the data again, it seems the images are too similar in each class. This would explain why the accuracy is so high on train, val AND test data - the data is simply to homogeneous. It might be interesting to test the model with another test dataset which has more image variance.
precision recall f1-score support
Apple Braeburn 1.00 0.80 0.89 50
Apple Granny Smith 1.00 0.98 0.99 50
Apricot 1.00 1.00 1.00 50
Avocado 1.00 1.00 1.00 44
Banana 1.00 1.00 1.00 49
Blueberry 1.00 1.00 1.00 47
Cactus fruit 0.94 1.00 0.97 49
Cantaloupe 1.00 1.00 1.00 50
Cherry 1.00 1.00 1.00 50
Clementine 1.00 1.00 1.00 49
Corn 1.00 1.00 1.00 45
Cucumber Ripe 0.91 1.00 0.95 40
Grape Blue 1.00 1.00 1.00 100
Kiwi 1.00 0.91 0.96 47
Lemon 1.00 1.00 1.00 50
Limes 0.98 1.00 0.99 49
Mango 1.00 1.00 1.00 49
Onion White 1.00 1.00 1.00 45
Orange 1.00 1.00 1.00 49
Papaya 1.00 1.00 1.00 50
Passion Fruit 1.00 1.00 1.00 49
Peach 0.81 1.00 0.89 50
Pear 0.99 0.99 0.99 70
Pepper Green 1.00 1.00 1.00 46
Pepper Red 0.99 1.00 0.99 67
Pineapple 1.00 1.00 1.00 49
Plum 1.00 1.00 1.00 46
Pomegranate 1.00 0.94 0.97 50
Potato Red 1.00 0.91 0.95 45
Raspberry 1.00 1.00 1.00 49
Strawberry 1.00 1.00 1.00 50
Tomato 0.99 1.00 0.99 75
Watermelon 1.00 1.00 1.00 48
accuracy 0.99 1706
macro avg 0.99 0.99 0.99 1706
weighted avg 0.99 0.99 0.99 1706
Classification report for simple model predictions.
Learning curves for simple model fit.
The finetuned VGG16 model reaches a perfect score of 100% acc after just 8 epochs. Being a larger, more complex and pretrained model, this was in every way expected. The model has weights from the ImageNet dataset training, and that dataset also includes fruits, which is beneficial for this project.
Compared to the simple model, the training time for this model was significantly longer. If the goal is to get high accuracy in predicting the test set labels, than it would be wiser to use the simple model.
precision recall f1-score support
Apple Braeburn 1.00 1.00 1.00 50
Apple Granny Smith 1.00 1.00 1.00 50
Apricot 1.00 1.00 1.00 50
Avocado 1.00 1.00 1.00 44
Banana 1.00 1.00 1.00 49
Blueberry 1.00 1.00 1.00 47
Cactus fruit 1.00 1.00 1.00 49
Cantaloupe 1.00 1.00 1.00 50
Cherry 1.00 1.00 1.00 50
Clementine 1.00 1.00 1.00 49
Corn 1.00 1.00 1.00 45
Cucumber Ripe 1.00 1.00 1.00 40
Grape Blue 1.00 1.00 1.00 100
Kiwi 1.00 1.00 1.00 47
Lemon 1.00 1.00 1.00 50
Limes 1.00 1.00 1.00 49
Mango 1.00 1.00 1.00 49
Onion White 1.00 1.00 1.00 45
Orange 1.00 1.00 1.00 49
Papaya 1.00 1.00 1.00 50
Passion Fruit 1.00 1.00 1.00 49
Peach 1.00 1.00 1.00 50
Pear 1.00 1.00 1.00 70
Pepper Green 1.00 1.00 1.00 46
Pepper Red 1.00 1.00 1.00 67
Pineapple 1.00 1.00 1.00 49
Plum 1.00 1.00 1.00 46
Pomegranate 1.00 1.00 1.00 50
Potato Red 1.00 1.00 1.00 45
Raspberry 1.00 1.00 1.00 49
Strawberry 1.00 1.00 1.00 50
Tomato 1.00 1.00 1.00 75
Watermelon 1.00 1.00 1.00 48
accuracy 1.00 1706
macro avg 1.00 1.00 1.00 1706
weighted avg 1.00 1.00 1.00 1706
Classification report for VGG16 model predictions.
Learning curves for VGG16 model fit.