Deep Learning-Based Nutritional Analysis of Food Images Using EfficientNetB4 and Multilayer Perceptron

Nutritional estimation is the growing research area within the domain of computer vision and deep learning for a balanced and healthy diet. Innovations in computer vision recently have shown greater improvements in food classification and analysis of nutritional values. Convolutional neural networks (CNNs) have resulted in a greater efficiency in the extraction of visual characteristics of food images. This paper proposes a fine-tuned EfficientNetB4 model along with multilayer perceptron (MLP), which resulted in an accuracy of 85% for food image classification. MLP regressor takes the feature vectors extracted from CNN backbone EfficientNetB4 for predicting nutritional values. The model is trained on Food-101 dataset with nutritional values metadata sourced from various sources to estimate nutritional targets. However, nutritional value estimation still presents a challenge due to mixed dishes and indefinite portion sizes. The dataset is preprocessed by normalization and resizing the food images, while nutritional labels are scaled using StandardScaler(). The model is also evaluated on various metrics such as mean absolute error and mean squared error and resulted in 0.30 mean absolute error and 0.21 mean squared error for estimating the nutritional values. Thus EfficientNet, complemented with an MLP, offers a robust framework for tackling the challenges posed by large-scale and varied datasets like Food-101 in the context of nutritional analysis. This hybrid model leverages the strengths of CNNs for fine-grained image feature extraction and the ability of MLPs to handle the final classification tasks, thereby enhancing the accuracy and reliability of predictions in real-world applications.