Combining deep residual neural network features with supervised machine learning algorithms to classify diverse food...

Area of Science:

• Computer Vision
• Machine Learning
• Health Informatics

Background:

• Obesity is a growing global health concern, linked to chronic diseases like type-2 diabetes and heart disease.
• Dietary monitoring via food logging is crucial for obesity prevention and management.
• Computer vision offers automated solutions for food logging through image classification.

Purpose of the Study:

• To evaluate the effectiveness of deep convolutional neural networks (CNNs) for automated food image classification.
• To assess the performance of pretrained ResNet-152 and GoogleNet models in extracting deep features for dietary intake monitoring.

Main Methods:

• Applied pretrained ResNet-152 and GoogleNet CNNs to extract deep features from diverse food image datasets (Food 5K, Food-11, RawFooT-DB, Food-101).
• Trained machine learning classifiers including artificial neural network (ANN), support vector machine (SVM), Random Forest, and Naive Bayes using extracted deep features.
• Validated model performance on multiple food image datasets.

Main Results:

• ResNet-152 deep features combined with SVM (RBF kernel) achieved 99.4% accuracy on the Food-5K validation dataset.
• ANN and SVM-RBF classifiers trained with ResNet-152 features demonstrated high accuracy on Food-11 (91.34%) and RawFooT-DB (99.28%) datasets.
• SVM with RBF kernel achieved 64.98% accuracy on the challenging Food-101 dataset.

Conclusions:

• Deep CNN features, particularly from ResNet-152, are highly effective for diverse food item image classification.
• Pretrained deep features offer strong generalization capabilities for automated dietary intake monitoring.
• This approach supports the development of advanced tools for healthy lifestyle maintenance and obesity management.