A Two-Color Space Input Parallel CNN Model for Food Image Classification

In recent years, machine learning approaches to food image classification have become popular because of the distinct characteristics of food classes, which exhibit low interclass variability and high intraclass variability. As the number of classes increases, these features become more challenging to extract, potentially making conventional machine learning techniques insufficient for robust and effective classification. This study introduces an alternative approach for classifying food images by utilizing different color spaces in a two-input parallel convolutional neural network (CNN) model. Using a pre-trained EfficientNetB0 as the underlying model for both subnetworks of the parallel CNN, the proposed approach has shown an increase in accuracy compared to the single CNN that only used RGB images as inputs. The proposed two-input parallel CNN model has achieved a Top-1 accuracy of 83.46% and a Top-5 accuracy of 96.34%. In addition, unlike the state-of-the-art results achieved by models that solely utilize RGB images as inputs, the two-input parallel CNN that accepts two color spaces as inputs attained higher accuracy using the same network architecture, specifically MobileNetV3Large and InceptionV3 as the underlying network. Hence, utilizing a two-input parallel convolutional neural network model enhances the model’s performance, regardless of the underlying network.