Abstract
With one in four individuals afflicted with malnutrition, computer vision may provide a way of introducing a new level of automation in the nutrition field to reliably monitor food and nutrient intake. In this study, we present a novel approach to modeling the link between color and vitamin A content using transmittance imaging of a pureed foods dilution series in a computer vision powered nutrient sensing system via a fine-tuned deep autoencoder network, which in this case was trained to predict the relative concentration of sweet potato purees. Experimental results show the deep autoencoder network can achieve an accuracy of 80% across beginner (6 month) and intermediate (8 month) commercially prepared pureed sweet potato samples. Prediction errors may be explained by fundamental differences in optical properties which are further discussed.
Introduction
Malnutrition remains a critical global health challenge, affecting approximately 25% of the population. This issue is particularly acute for individuals suffering from dysphagia (difficulty swallowing), who often rely on pureed diets for safe ingestion. However, the nutritional quality of pureed food is notoriously variable, depending heavily on preparation methods and dilution. Currently, monitoring nutrient intake in clinical or home settings relies on time-intensive manual checks, which are prone to subjectivity and error.
This research aims to bridge the gap between visual appearance and nutritional content by targeting Vitamin A (carotenoids). Because carotenoids are optically active in the visible spectrum, they serve as ideal candidates for computer vision-based assessment. By quantifying the relationship between color and molecular concentration, the authors propose a system capable of objective, automated nutrition monitoring.
Methodology
The proposed system utilizes a "two-pronged" approach combining physics-based biophotonic analysis with deep learning techniques.
Biophotonic Analysis
The system employs transmittance imaging, where light from a broadband tungsten-halogen source passes through a sample placed in a petri dish. The resulting images are normalized and mapped to the CIELAB color space. This color space is particularly effective because its coordinates correlate closely with the chemical structure of carotenoids—specifically the conjugated double bonds that determine light absorption.
Deep Learning and Physics Modeling
To estimate bulk nutrient density, the researchers implemented a deep autoencoder network. This model, originally trained on reflectance data, was fine-tuned using the new transmittance dataset. The network performs a "top-1 class prediction" to identify the relative concentration (dilution level) of the food sample. This is supported by a photon migration model based on the Beer-Lambert law, which accounts for how water and Vitamin A attenuate light within the food matrix.
Experimental Setup
The study utilized five commercially prepared pureed foods known for high Vitamin A content: butternut squash, carrot, mango, and two types of sweet potato (6-month "beginner" and 8-month "intermediate"). To simulate varying nutrient densities, a five-tier dilution series was created for each food type. The final dataset for the deep learning component consisted of 400 RGB samples extracted from the transmittance images.
Results and Discussion
Prediction Accuracy
The fine-tuned deep autoencoder demonstrated strong performance in identifying dilution levels. For sweet potato purees, the system achieved a maximum top-1 prediction accuracy of 80%. When analyzed individually, the model reached 79% accuracy for 6-month samples and 81% for 8-month samples, demonstrating consistency across different preparation textures.
Optical Insights and the "Carrot Paradox"
The study confirmed that color serves as a reliable chemical proxy; in the CIELAB space, higher Vitamin A concentrations move toward the origin, while lower concentrations shift upward and to the right. However, the researchers noted a "carrot paradox": while sweet potato had the highest total Vitamin A content, carrot appeared visually "redder." This discrepancy is likely due to the specific ratio of $\beta$- and $\alpha$-carotene. Carrots may contain a higher proportion of specific absorbers that transmit red light more effectively despite having lower overall carotenoid levels.
Furthermore, the system detected subtle optical differences between 6-month and 8-month baby foods that are not typically reflected on nutritional labels. The 6-month sweet potato was found to be redder and lighter, suggesting that the system can distinguish between preparation methods and processing stages.
Conclusion
While the current results are promising, the study identifies several areas for future refinement. The current model primarily assumes Vitamin A is the dominant absorber, but other constituents like chlorophyll, iron, and starches cause light scattering that can interfere with the signal. Additionally, Vitamin A is sensitive to heat and oxidation, meaning the "actual" nutrient value in a processed puree may differ from the raw ingredients.
Future work will focus on expanding the dataset to include a wider variety of nutrients and food types. By integrating data fusion techniques, the researchers hope to better distinguish between absorption (nutrients) and scattering (texture and fats), moving closer to a holistic, automated nutrient assessment tool.
Summary updated