Recognising classic Hong Kong style dishes using Transfer Learning with Residual Neural Network

Developed By: Arjun Rao and Mudit Chaudhary - The Chinese University of Hong Kong

Dataset Provided :

• Food images classified into 75 "large categories" - each containing up to 1000 images each category.
• 25 small categories - containing 5 images per category

Solution used - Deep Residual Network with Transfer Learning (ResNet-50)

Data augmenation was used to enhance

• Random rotations with range 40 degrees

• Height and width shift

• Zoom

• Shear

• Horizontal flip

• Training Methodology:

1. Train the ResNet-50 with images from the large sample
2. Freeze the weights of that model and implement a new output layer for training the images from small sample.
3. Train last few layers of the previous model.
4. Transfer Learning from previous models

Design methodology

For the first model: We add a Global Average Pooling Layer connecting ResNet. Add a Fully Connected layer with 1024 nodes Add a Dropout Layer with 40% dropout rate Add a Fully Connected output layer with 75 nodes For the second model: We add a Global Average Pooling Layer connecting ResNet. Add a Fully Connected layer with 1024 nodes Add a Dropout Layer with 40% dropout rate Add a Fully Connected output layer with 25 nodes

Future improvements :

Future plan include: Tuning the hyperparameters Using Inception ResNet Modify the output layer to reduce overfitting for the small sample