The study developed a two-stage ThaiSook pipeline: a binary model first rejects non-produce images, after which a multiclass model identifies 43 fruit and vegetable types and three vegetable-heavy dishes. The 40,595-image dataset contained 11,210 labelled and 29,385 unlabelled user images. With label smoothing and temperature scaling, the pipeline reached 92.43% top-1, 94.56% top-3, and 94.91% top-5 accuracy; Noisy Student raised top-1 accuracy to 92.59%. However, results rely on one train–test split, without real on-device testing or evidence that classification improves dietary measurement or health outcomes.
Key findings
- The two-stage model with label smoothing α=0.10 and temperature scaling T=2.00 achieved 92.43% top-1, 94.56% top-3, and 94.91% top-5 accuracy. Adding Noisy Student produced 92.59% top-1 accuracy versus 82.11% for the one-stage baseline—a gain of 10.48 percentage points.
- EfficientNetV2S was the strongest of seven compared binary backbones, reaching 95.62% accuracy, 95.61% F1, 95.57% AUROC, and 96.82% PRAUC. In multiclass experiments, however, Noisy Student added only a small end-to-end improvement over the already tuned label-smoothing and temperature-scaling model.
- Class performance was uneven: weighted F1 was 0.9250 while macro F1 was 0.8988. Banana, broccoli, custard apple, jackfruit, and strawberry exceeded 0.95 F1, whereas cantaloupe reached 0.6400, musk melon 0.7500, and Somtam 0.7692. Similar colour and shape, mixed ingredients, background, illumination, and image quality drove failures.
- On Fruits-360, α=0.10 and T=2.00 produced 98.53% accuracy, 98.45% AUROC, and 99.20% F1. This is a separate public benchmark, not external validation on new ThaiSook users or under field-deployment conditions.
Why this matters globally
A classifier covering Southeast Asian produce and local vegetable dishes could reduce manual food-log review and enable faster feedback, particularly for foods underrepresented in global datasets. Its broader contribution is a context-specific development strategy combining a modest labelled dataset with unlabelled user images. Category accuracy, however, is not equivalent to accurate portion-size, nutrient-intake, or health-outcome measurement. Deployment requires testing on genuinely new users across countries, devices, lighting conditions, and mixed meals, followed by evaluation of whether the system improves dietary data quality or behaviour.
Thai researcher contribution
NSTDA developed the ThaiSook platform and is the affiliation of Decho Surangsrirat and Panyawut Sri-iesaranusorn. Polathep Vichitkunakorn is affiliated with three units at Prince of Songkla University’s Faculty of Medicine and is the corresponding author. The publisher’s CRediT statement assigns all three Thai-affiliated authors broad roles spanning conceptualisation, methodology, software, data curation and validation, formal analysis, writing, project administration, and supervision. ThaiSook user data and ethics approvals from NSTDA and Prince of Songkla University place Thai institutions at the centre of the data, system, and governance rather than as nominal collaborators.
Limitations to consider
The main estimate comes from a single train–test split, without repeated cross-validation or confidence intervals. The paper does not describe a separate validation partition despite early stopping on validation loss, and architecture and parameter selection within the same pipeline may yield optimistic performance. The dataset is locally biased toward Thai produce and foods; no public download link for the full image dataset is supplied, and pseudo-labels can propagate teacher-model errors. Temperature scaling was used without Expected Calibration Error or reliability diagrams. The reported test counts of 1,168 and 1,091 images do not align exactly with 20% of the table totals, and no explanation is given. Crucially, the 83.16 MB model size, 0.1131-second inference time, 8.84 FPS, and 1.83 GB RAM use were measured in a server-grade/Google Colab Tesla T4 environment, not on a phone. TensorFlow Lite conversion, quantisation, and on-device latency remain future work.