Singaporean Food Image Classification

Objective

Build a food image recognition technology to recognize more than 10 Singaporean food. Model can then be reused to develop a food recommendation website or app based on the image uploaded by the users.

Accomplishments

• Dataset collection: we utilised data scrapers to crawl images from various sources and saved the images, path and category to a csv file.
• EDA: We went thrrought multiple checks to clean our dataset, by removing duplicates, removing images with multiple food items and many more to get a clean dataset.
• Data Augmentation and Preprocessing: Resize, Normalization, Color Saturation, Random flip were some techniques we used to increase our dataset and improve the quality of it.
• Model Training: We implemented MLP as our baseline model initially, then implement CNN.
• Model Improvement: Since both the models did not perform very well, we introduction batch normalization for MLP and ResNet instead.

Results

ResNet with transfer learning(from ImageNet weights) had the highest accuracy of 0.77.