CNN Based Study of Improvised Food Image Classification

Food recognition has captured numerous research attention for health-related applications. Food recognition is a challenging task due to the diversity of food. Convolutional Neural Networks have addressed the complex feature extraction problem and it has improved the classification accuracy compared to traditional image processing techniques. There are different ways to build the food classification model (i.e.) building CNN from scratch, transfer learning, one shot learning, iterative learning and so on. Transfer learning helps in a better way in generic feature extraction and improves classification accuracy. Along with transfer learning, data augmentation techniques have improved the overall classification summary. In this paper, an automated augmentation technique to remove background objects in food images along with an additional augmentation technique is tried. This has improved the classification accuracy and also partial dataset has been used and compared the classification accuracy for different data set sizes. When the background objects in the food images are removed, CNN trains faster and also provides better performance.