Utilizing Color Space Information as Features for Deep Learning for a Heterogeneous Food Recognition System

Food recognition serves as one of the most promising applications of visual object recognition, as it can be used to calculate food calories and analyze people's eating habits and dietary practices for health-care purposes. It forms the primary and most vital step in developing an application capable of providing nutritional recommendations. In this work, the CNN, ResNet, and AlexNet architectures have been proposed, given their popularity among researchers. A merged dataset, comprising Food-101 and UEC-Food256, was employed to evaluate the proposed models. To ascertain the optimal color spaces for food recognition, eight selection criteria were proposed. Among the various color space selection criteria, RGB showed superior performance when used with the Custom CNN model. To enhance performance, an upsampling method was implemented to increase the number of samples in the minority classes. This was achieved by duplicating existing samples using an augmentation process, thereby balancing the merged dataset. The final results suggest that EfficientNetB0, a CNN-based pre-trained model, performs better with the RGB color space, increasing accuracy from 40.48% (Custom CNN) to 73.11% (EfficientNetB0).