SCALEeat: Vision-Guided Food Scale for Automated Macronutrient Estimation ^†

Abstract

SCALEeat, a self-contained smart food scale, was developed to offer a convenient solution and replace manual logging with on-device recognition and weighing. The device integrated a Raspberry Pi 5, a camera, and a load cell, identifies foods and computes calories, carbohydrates, protein, and fat from measured weight through the Philippine Food Composition Tables (PhilFCT). Using transfer learning, a MobileNetV3-Large model trained on 25 commonly consumed items from ENNS, this achieved a 97.33% top-1 accuracy on a 300-image test set. Deployed on the prototype, SCALEeat achieved 93.60% accuracy, demonstrating practical accuracy and a lower-friction path to routine dietary assessment.

1. Introduction

Technology-enabled self-monitoring has heightened interest in smart tools for meal planning and dietary tracking. However, leading nutrition applications frequently misestimate intake—underreporting energy and lipids and inconsistently estimating protein—primarily because they rely on non-localized composition tables and user-generated entries [1]. Adoption is further constrained by usability and trust concerns, including database quality and the effort required for manual input, underscoring the need for tailored and reliable solutions [2].

Ingredient recognition has proven feasible: edge vision systems report 85–90% accuracy; an 84.76% precision and recall of You Only Look Once Version 8 (YOLOv8) with cosine similarity, and 90% of MobileNetV2-based RecipeLens. This shows reliable detection for downstream nutrition tasks [3,4]. In contrast, image-only calorie estimation remains error-prone. A chamber-based CNN with an ultrasonic height sensor reported a 12% calorie error [5], while single-image “chopstick reference” methods achieved 6–7% weight/calorie error [6]. Both approaches infer mass from volume proxies rather than directly weighing portions, allowing bias to accumulate. This limitation motivates the development of systems that integrate recognition with physical weighing to improve nutrient estimation reliability.

In this study, we developed SCALEeat, a smart food scale that performs on-device food identification and macronutrient estimation at the point of preparation, thereby reducing manual logging and aligning outputs with locally relevant foods. The system integrates a Raspberry Pi 5, camera, and YZC-131 load cell with a MobileNetV3-Large classifier adapted via transfer learning to an Expanded National Nutrition Survey (ENNS)-based dataset. Nutrient values are derived from the Philippine Food Composition Tables (PhilFCT) per 100 g and are scaled by the measured mass to compute calories, carbohydrates, protein, and fat. We used a set of 25 commonly consumed foods, detailing the dataset, model design, and hardware workflow, and evaluated system performance on held-out tests.

2. Materials and Methods

2.1. System Workflow

The scale’s process begins as the user places a food item on the scale; the camera captures an image while the load cell records net weight. The image is preprocessed and classified by the onboard CNN. The predicted label pulls nutrient values from the local PhilFCT database, which are scaled by the measured mass to compute calories, carbohydrates, protein, and fat, then displayed to the user (Figure 1).

2.2. Data Preparation

2.2.1. Nutritional Database and Food Classes

We trained on 25 classes drawn from ENNS’ most-consumed Filipino foods [7]. Per-100 g energy and macronutrient values (kcal, carbohydrates, protein, fat) were obtained from the Philippine Food Composition Tables (PhilFCT,

Table 1. Nutritional database and food classes per 100 g.

No	Class Name	kcal	C	P	F	Number	Class Name	kcal	C	P	F
1	Banana—unpeeled	126	29.6	1.4	0.2	13	Instant noodles—uncooked	429	57.3	10.5	17.5
2	Bread loaf	329	61.1	9.7	5.1	14	Mango ripe—unpeeled	70	16.4	0.6	0.2
3	Cabbage—uncooked	31	5.9	1.6	0.1	15	Mayonnaise	715	1.4	1.2	78.3
4	Carrot—uncooked	42	8.6	1.1	0.3	16	Pandesal	330	62.9	10.1	4.2
5	Chicken white meat—boiled	173	0.0	32.9	4.6	17	Pork kasim—fried	136	0.0	20.5	6.0
6	Chicken white meat—Fried	226	0.0	40.3	7.2	18	Potato—uncooked	78	16.8	2.4	0.1
7	Cooking Oil	896	0.0	0.0	99.6	19	Powdered milk	148	11.7	7.7	7.8
8	Egg—fresh	139	1.4	12.3	9.4	20	Red onion—unpeeled	52	10.5	1.7	0.3
9	Fuji apple	65	15.6	0.2	0.2	21	Salted egg—red	192	4.4	13.6	13.3
10	Garlic—unpeeled	129	24.6	7.0	0.3	22	Sardines in tomato sauce	88	2.7	9.7	4.3
11	Hotdog—uncooked	226	4.8	13.4	17.0	23	Spanish bread	371	55.5	9.3	12.4
12	Indian mango—unpeeled	53	12.5	0.3	0.2	24	Tomato—uncooked	25	5.2	0.8	0.1
C = carbohydrates, P = protein, F = fats per 100 g.						25	White rice—boiled	129	29.7	2.1	0.2

) [8]. To account for preparation effects, certain items were divided into subclasses (e.g., chicken: boiled or fried). Labels include state qualifiers (e.g., unpeeled, boiled), and when no preparation method is specified, the item is treated as uncooked.

2.2.2. Image Dataset and Augmentation

The image dataset comprised 50% researcher-captured photos and 50% open-source images for diversity. Per class, we allocated 35, 15, and 12 images to training, validating, and testing datasets. As illustrated in Figure 2, training images were resized and padded to 640 × 640 (black fill) and augmented in Roboflow (version 3.0, Roboflow Inc., Des Moines, IA, USA), expanding the training set to 8750 samples (Figure 2). Validation and test sets were unaugmented to enable unbiased performance estimation.

2.3. Classification Model

2.3.1. Model Architecture

The classifier uses MobileNetV3-Large for high accuracy-per-compute on embedded hardware [8]. It employs a flexible input with resize to 320 × 320, and a lightweight head (Dense + Sigmoid Linear Unit; final Softmax). It includes batch normalization and two dropout layers (0.20, 0.10) to limit overfitting. With ~3.25 M parameters, the model runs on a Raspberry Pi in real time while maintaining strong classification accuracy (

Table 2. Model layers and parameters.

Layer	Output Shape	Activation Function	Number of Parameters
Flexible input (input layer)	(None, None, 3)	-	-
resizing layer	(320, 320, 3)	-	-
MobileNetV3Large (pre-trained)	(None, 960)	-	2,996,352
Batch norm 1	(None, 960)	-	3840
Dropout 1	(None, 960)	-	-
Dense 1	(None, 256)	SiLu	245,760
Batch norm 2	(None, 256)	-	1024
Dropout 2	(None, 256)	-	-
Dense4 (classifier)	(None, 25)	Softmax	5418

).

2.3.2. Model Training

In training, we used two-stage transfer learning with Adam and categorical cross-entropy. In Stage 1, the custom head is trained with the MobileNetV3 backbone frozen, while in Stage 2, the model is fine-tuned at a reduced learning rate, unfreezing all layers except batch-norm to adapt features to the dataset. Training is stopped on validation loss to prevent overfitting.

2.4. Hardware Development

The device with the model integrates a Raspberry Pi 5 with a 7-inch touchscreen, a camera with a white LED, and a 5 kg YZC-131 load cell read by an HX711 24-bit ADC in a reinforced 3D-printed enclosure (Figure 3). The load cell was calibrated in two steps: (1) zero offset from ~30 empty readings; (2) scale factor from a known mass to convert counts to grams.

2.5. Performance Evaluation

The evaluation was conducted using a confusion matrix along with five key metrics. These metrics included accuracy, which measures overall correctness; precision, which reflects the proportion of correctly identified positives; recall, which indicates the proportion of actual positives that were successfully recovered; the false-positive rate, which captures the proportion of incorrect positives; and the F1-score, which represents the harmonic mean of precision and recall.

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{C}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{S}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}},\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)}},$$

(1)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})}},\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e} \mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e} \mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})}},\mathrm{F}1-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2\times {\displaystyle \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{(\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l})}}$$

(2)

3. Results and Discussion

3.1. Model Training and Validation

We trained a MobileNetV3-Large classifier via transfer learning using Adam and categorical cross-entropy with a batch size of 32 and the optimized hyperparameters presented in

Table 3. Optimized model hyperparameters.

Hyperparameter	Value	Description
INPUT_SIZE	320 × 320	Training set input image dimensions
EPOCHS_STAGE1	20	Max epochs for training the classifier head
EPOCHS_STAGE2	80	Max epochs for fine-tuning the model
LR_STAGE1	10−3	Learning rate for the first training stage
LR_STAGE2	10−4	Learning rate for the fine-tuning stage
WEIGHT_DECAY	10−4	Regularization parameter to prevent overfitting
LABEL_SMOOTHING	0.1	Regularization technique for the loss function
MONITOR	val_loss	Metric used for early stopping
PATIENCE	10	Epochs to wait for improvement before stopping

.

Figure 4 illustrates the validation accuracy and loss across epochs during model training. Stage 1, with the backbone frozen, exhibited an early plateau, reaching 95.47% validation accuracy at approximately epoch 17. Subsequent fine-tuning in Stage 2 surpassed this threshold, increasing validation accuracy to 97.60% while reducing validation loss, indicative of improved generalization. Performance then stabilized, confirming effective adaptation of pretrained features to the food dataset. Although the peak validation accuracy reached 97.60%, the corresponding checkpoint was associated with a higher validation loss (0.7946). In practice, the model achieving 97.07% accuracy with a lower validation loss (0.7305) demonstrated superior generalization and was therefore selected for deployment. This outcome underscores that validation loss, rather than marginal gains in accuracy, often provides a more reliable criterion for model selection and generalization.

3.2. Performance on the Test Set

The selected model was evaluated on a held-out test set comprising 300 images independent of the training data.

Table 4. Overall model performance on the held-out test set.

Metric	Value
Accuracy	0.9733
Precision (macro)	0.9741
Recall (macro)	0.9733
F1-score (macro)	0.9737
False positive rate (macro)	0.0011

presents the performance metrics, computed using functions from the scikit-learn library. The model achieved an overall accuracy of 97.33%, demonstrating strong classification capability for food items. With a macro-averaged precision of 0.9741 and an F1-score of 0.9737, the model effectively balanced correct identifications while minimizing misclassifications. Furthermore, its low false-positive rate of 0.0011 highlights the robustness and reliability of its predictions.

3.3. Deployed System Performance and Discussion

When the model predicts chicken white meat–fried, the device looks up its per-100 g nutrients in PhilFCT (226 kcal, 40.3 g protein, 0.0 g carbs, and 7.2 g fat) and then scales each value by the measured weight. For a portion weighing 81.11 g, the system multiplies 0.8111 by each per-100 g value, yielding ≈183 kcal, ≈32.7 g protein, 0.0 g carbohydrates, and ≈5.8 g fat (rounded to the display precision). For testing, we used 3–5 exemplars per class and created ten trials per item by varying background, orientation, cutting, and portion size (Figure 5).

From the actual testing, the system achieves 234 correct predictions out of 250 samples (10 samples per class), which gives an overall accuracy of 93.60% in a real-world setting still validates the system’s effectiveness, even though it drops from 97.33% accuracy in the test set. Using the calculated true positive (TP), true negative (TN), false positive (FP), and false negative (FN) per class on the testing and averaging the performance per class gives us the macro performance of the model, as shown in

Table 5. Overall model performance in real-world testing.

Metric	Value
Accuracy	0.9360
Precision	0.9437
Recall	0.9360
F1-score	0.9398
False positive rate	0.0027

. The precision, recall, and F1-score values align with the accuracy value, showing a good model performance in all the metrics. Given the environmental variance, a slight increase in the false positive rate of 0.0027 is expected, yet it remains deficient.

4. Conclusions and Recommendations

The MobileNetV3 model achieved 97.33% top-1 accuracy on a 300-image test set and 93.60% accuracy when integrated into SCALEeat, the end-to-end prototype, establishing a strong baseline for a vision-guided food scale with automated macronutrient estimation. The system reliably identifies single items and logs kcal, carbohydrates, protein, and fat from measured weight using PhilFCT. For future work, expand the class set and diversify the dataset (more images per class across backgrounds, lighting, preparation states, and regions) to improve dietary coverage.