Image Classification with Transfer Learning

Transfer learning is repurposing a pre-trained model for another but similar usage. This method is seen in various machine learning applications especially in situations where the dataset is relatively small.

In this project, I built an image classification model from scratch using transfer learning. When I said from scratch, the dataset was my custom dataset which I scrapped using the IDT tool with my own custom classes. You can read about this tool here.

Problem statement:

Most people tend to have issues classifying bags and carry-ons. In this project, I chose to work with 5 classes namely; Backpack, Briefcase, Duffle, Handbag and purse.

Data collection and cleaning:

Like I stated earlier, I used the IDT tool to scrape images from the web then I cleaned the data which happened to be the most tedious of all processes. In all, the dataset contained 2000 images divided into 400 images per class which I cleaned manually. I removed duplicates and images with associated keywords, but has no correlation with the intended images.

Data preprocessing:

I rescaled the images and performed a validation split of 20% before performing the image data generator function from Keras.

Then I got the class labels and indices.

Model building:

Using Tensorflow hub, I imported a pre-trained model which was the second version of mobilenet because of its scalability.

I compiled the model using Adam optimizer with learning rate of 0.001, categorical_crossentropy loss and the ‘metrics’ accuracy.

Model training:

I trained the model using 10 epochs.

Then I used matplotlib to plot the graph of the training of the model.

Model Prediction:

After I viewed the validation batch shape and prediction result shape, I visualized the model’s predictions with the correct labels in green and the incorrect labels in red.

Conclusion:

In the project folder, there are two more models done without transfer learning and looking at their graphs, they were overfitting. The project is found here and I would welcome feedbacks on this project. You can connect with me on LinkedIn. Thank you for reading