Toyota_Model_Recognizer

This Web App uses the EfficientNet_Lite0 model to build an API that predicts the model of toyota vehicles.

Steps

The scrape_car_images_using_selenium.ipynb notebook was executed locally to scrape one thousand, one hundred (1100) images of toyota vehicles from google. The scrapper only scrapped for 10 vehicle classes, namely:

• toyota_4runner
• toyota_avalon
• toyota_camry
• toyota_corolla
• toyota_fjcruiser
• toyota_hiace
• toyota_hilux
• toyota_landcruiser
• toyota_rav4
• toyota_sienna

These classes were selected on the basis of brand model popularity within the Nigerian context. The selection followed no particular hierarchy. All selected vehicle classes had 110 images each in a class-separated folder.

Link to the dataset can be found here

DataLoader

The Image_Classification_with_TensorFlow_Lite_Model_Maker.ipynb notebook was executed in google colab for efficeincy and speed. The images from the folders were loaded using the DataLoader function from tflite_model_maker.image_classifier class. The DataLoader function together with the from_folder method was used to load images from subdirectories, identifying the subdirectory names as the class labels.

Model Training

The default EfficientNet_Lite0 model from tflite_model_maker.image_classifier.image_classifier.create class was used to train the images for 50 epochs achieving an accuracy score of 72.7% on the validation dataset. This is considered a good score considering the very small volume of data at the model's disposal.

Model Compression

A pre-trained VGG19 model was first used to train the images as seen in the keras_vgg19_toyota_model_recognizer.ipynb notebook. The top layer was removed to accomodate for a fully connected Dense Layer with 1024 neurons, a relu activation function, batchnormalization and dropout layers. Softmax activation function was applied at the output layer for multiclass classification (in this case 10 classes). The size of the model after training was 347mb. This large file size caused redundancies in deployment, as github does not permit pushing of files greater than 100mb.
For this reason the model was compressed by following the steps outlined in this official tensorflow doc for optimization and quantization of models. This helped reduce the model size x10, to about 43mb, while maintaining the model's predictive accuracy.
The compress_keras_model.ipynb notebook shows the code for this. The backlog of this compression was that the compressed model took about 1 min 20 seconds to perform a single prediction, which resulted in very bad user experience.

To further reduce the model size the tflite_model_maker package was used to build a .tflite model and achieved an even better efficiency (model accuracy of 0.727 and model portability of 3.8mb) compared to the baseline model which had an accuracy score of 0.70 with a size of 370.7 mb and the compressed model which had an accuracy score of 0.67 with a size of 47mb.

The tflite_model_maker model is stored in the model_dir folder as model.tflite.
The fully trained VGG19 model together with the compressed model is found here

Running the service...

• Click here to run a prediction
• Click on Choose File to select an image
• Select a toyota image from your local machine
• Click on the Predict button to the right
• Wait a few seconds for the system to process and predict the image class.
• Viola!!! Here you have your prediction.

Snapshot of Web App

(The home page)

The prediction Page

Further Steps

• Create a more sophisticated web interface
• scrape more image dataset to create an even better model
• Train model using Transfer learning and compare performance
• Include more classes

PS: The Image_Classification_with_TensorFlow_Lite_Model_Maker.ipynb notebook was executed with the help of this tensorflow tutorial