Vision-Based Convolutional Neural Network System for Automated Recognition of Infant Complementary Foods

Automatic introduction of complementary foods plays an important role in supporting objective and efficient nutritional monitoring of young children. However, the visual characteristics of complementary foods, which tend to be homogeneous in terms of color, texture, and shape pose significant challenges for computer vision-based recognition systems. This study aims to develop and evaluate a deep learning-based system for automatic recognition of complementary foods. This study proposes an image classification system for complementary foods based on Convolutional Neural Network (CNN) using ComFoodID21, a specialized dataset designed to represent the visual characteristics of complementary foods. The main contributions of this research include the development of a CNN-based recognition system for homogeneous visual domains, the compilation of ComFoodID21 as an initial benchmark dataset, and the evaluation of multiple CNN architectures in this domain. Three CNN architectures, namely EfficientNetB0, ResNet50, and MobileNetV2, were initialized using ImageNet pre-trained weights and fine-tuned on the ComFoodID21 dataset. Performance was evaluated using accuracy, precision, recall, F1-score, and training time analysis to assess computational efficiency. Experimental results show that ResNet50 provides the best trade-off between accuracy and training efficiency, achieving 98.28% accuracy with faster convergence despite a slightly longer per-epoch computation time. EfficientNetB0 attains comparable accuracy but requires more epochs to converge, while MobileNetV2 yields lower accuracy and slower convergence in the homogeneous visual domain. The proposed system demonstrates potential for application in smart nutrition monitoring and decision-support systems. The ComFoodID21 dataset is available to the research community and can be accessed at https://tinyurl.com/datasetcomfoodid21.