Improved food recognition using a refined ResNet50 architecture with improved fully connected layers

Food consumption has significant effects on human health, particularly in relation to quality control, quantity, freshness, and color. This study focuses on identifying food across 16 categories, divided into breakfast, lunch, and dinner, to evaluate its impact on the human body, specifically in hospital and restaurant settings. The recognition system was used a machine vision system and deep learning algorithms to record food consumption videos, extracting images. After preprocessing, a raw dataset was bulit that consisted of 12,000 images, expanded to 66,000 images through data augmentation. Five deep learning algorithms were used for recognizing food and consumed food. ResNet50 was the best algorithm in comparison to other deep learning architectures. The effect of Hyper-parameters such as data augmentation, batch size, image size, and learning rate on performance of Resnet50 were analyzed. Transfer learning method led us to develop three versions: standard ResNet50, fine-tuned ResNet50, and optimized ResNet50 with a customized fully connected layer. ResNet50 with a specific dense layer was the best development version of ResNet50. This model with Adam optimizer, 10^-3 initial learning rate, batch size 4, and image size 340 × 640 could recognize various foods with 97.25% accuracy and 0.2 loss. Response time and training time of this architecture compared to other algorithms were confidential; the training process and response time were 5.30 h and 1.2 s. ResNet50 with a specific fully connected layer powerfully could complete tasks with high accuracy and the least time.