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. We present results on the link between color and vitamin A content using transmittance imaging of a pureed foods dilution series in a computer-vision powered intelligent nutrient sensing system prototype and use a fine-tuned deep autoencoder network to predict the relative concentration of sweet potato purees. Results indicate an network accuracy of 80% across beginner (6 month) and intermediate (8 month) commercially prepared pureed sweet potato samples. Network errors may be explained by fundamental differences in optical properties which are further discussed.
Introduction
Malnutrition remains a global health crisis, affecting approximately 25% of the population. This issue is particularly acute for the 590 million individuals worldwide suffering from dysphagia (swallowing difficulties). To ensure safe ingestion, these patients rely on pureed foods. However, the nutritional density of these purees varies significantly based on preparation methods and commercial manufacturing. Currently, monitoring nutrient intake requires labor-intensive manual assessment. This research proposes an automated computer vision system to objectively estimate nutrient density, specifically Vitamin A, in homogeneous pureed foods.
System Methodology
The proposed system utilizes a two-pronged approach to nutritional assessment: biophotonic analysis for single nutrient assessment and a deep learning framework for bulk relative nutrient density prediction.
Biophotonic Analysis
Unlike standard reflectance photography, this system employs transmittance imaging. A broadband tungsten-halogen light is passed through a 15 mL food sample. The resulting images are normalized and converted into the CIELAB color space. This modality is chosen because Vitamin A (carotenoids) acts as a chromophore, absorbing specific wavelengths. By modeling light attenuation via the Beer-Lambert law, the system can correlate color shifts with nutrient concentration.
Deep Autoencoder Network
For coarse-grained dilution estimation, the researchers implemented a deep autoencoder network. The model architecture consists of two primary autoencoder stages (reducing dimensions from 3750 to 100, then to 50) followed by a softmax activation layer. The network was fine-tuned using 400 RGB samples extracted from transmittance images to predict five distinct dilution levels (20% to 100%).
Experimental Setup
The study utilized a custom dataset of five commercially prepared pureed foods: butternut squash, carrot, mango, and two varieties of sweet potato (6-month "beginner" and 8-month "intermediate"). For each food type, a five-tier dilution series was created to simulate varying nutrient densities. The final dataset for the autoencoder included 400 RGB samples (25x50 pixels) derived from 30 full-field transmittance images.
Experimental Results
The fine-tuned deep autoencoder demonstrated robust performance in predicting nutrient concentration levels. When tested on a combined set of 6-month (sp6) and 8-month (sp8) sweet potato samples, the system achieved a maximum top-1 prediction accuracy of 80%.
As shown in the performance table, the model exhibited higher precision when classifying the 8-month sweet potato samples individually, reaching a maximum accuracy of 92%. The 6-month samples peaked at 79% accuracy. These results suggest that while the network is effective at generalizing across similar food types, inherent optical differences between product formulations affect prediction reliability.
Discussion and Optical Complexity
A key finding of the study is the complexity of the color-nutrient link. For instance, while carrots appeared "redder" in the CIELAB space, they possessed a lower Vitamin A % Daily Value than sweet potatoes. This is attributed to the specific ratio of $\beta$-carotene to $\alpha$-carotene, which alters the absorption profile in the blue-green spectrum. Furthermore, the system identified distinct optical "fingerprints" for 6-month versus 8-month sweet potato purees, indicating that manufacturing processes introduce nutritional variations not always apparent on standard labels.
Conclusion and Future Work
This work demonstrates the feasibility of using transmittance imaging and deep learning to automate the nutritional assessment of pureed foods. Future iterations aim to move beyond the simplified assumption that Vitamin A is the sole absorber by accounting for interference from chlorophyll, iron, and light-scattering constituents like fats and starches. Expanding the dataset to include a wider variety of nutrients and food types will be essential for a holistic clinical tool.
Summary updated