Automated food recognition is essential in order to streamline dietary monitoring. To build and evaluate complex food recognition models, large datasets of annotated food images are crucial. In this paper, we introduce a new dataset called THFOOD-100, which is specifically designed for this purpose. This dataset consists of 53,459 high-quality images of popular Thai dishes categorized into 100 classes. We conducted a comprehensive comparison of 23 deep convolutional neural network and vision transformer architectures to establish a strong baseline for classification performance on the THFOOD-100 dataset. Additionally, we proposed training the models using cyclical learning rates, which has been shown to improve model generalization and significantly reduce training time. We demonstrated the effectiveness of cyclical learning rates with three standard optimizers on THFOOD-100, ETHZ Food-101, and UEC-Food256. The top-performing model achieved a 96% classification accuracy on THFOOD-100, showing great promise for real-world applications. Our new dataset is specifically aimed at better representing Thai cuisine in food recognition research, and our analyses offer valuable insights into the shortcomings of current models.

Obesity and diabetes represent escalating health challenges, exacerbated by the widespread availability and convenience of processed foods. Automated dietary monitoring technologies offer valuable support to individuals managing these conditions by helping them track and regulate their consumption, particularly of high-calorie, sugar-rich foods. Toward this goal, this paper presents KediNet, a hybrid deep learning architecture that combines the strengths of convolutional neural networks (CNNs) and vision transformers (ViTs) for Thai dessert recognition. Our main contributions include: 1) the proposal of KediNet, which outperforms nine state-of-the-art deep learning models; 2) a comprehensive comparison of classification performance across different model architectures, image resolutions, and data augmentation techniques; and 3) the creation of THDESS-20, a new publicly available dataset containing 5,242 images across 20 classes of Thai desserts. Experimental results demonstrate that KediNet achieves 98.64% balanced accuracy at $224\times 224$ resolution, surpassing all other tested models even at higher resolutions, while maintaining inference times suitable for mobile applications. Our analysis shows that KediNet effectively extracts meaningful features from Thai dessert images despite their high intra-class variation and visual similarities between classes. This work advances the field of food recognition for health applications while addressing the specific challenges of ethnic cuisine classification.

Food ingredient prediction from images is a challenging multi-label classification task with significant applications in dietary assessment and automated recipe recommendation systems. This task is particularly difficult due to highly imbalanced classes in real-world datasets, where most ingredients appear infrequently while several common ingredients dominate. In such imbalanced scenarios, the F-score metric is often used to provide a balanced evaluation measure. However, existing methods for training artificial neural networks to directly optimize for the F-score typically rely on computationally expensive hyperparameter optimization. This paper presents a novel approach for direct F-score optimization by reformulating the problem as cost-sensitive classifier optimization. We propose a computationally efficient algorithm for estimating the optimal relative cost parameters. When evaluated on the Recipe1M dataset, our approach achieved a micro F1 score of 0.5616. This represents a substantial improvement from the state-of-the-art method’s score of 0.4927. Our F-score optimization framework offers a principled and generalizable solution to class imbalance problems. It can be extended to other imbalanced binary and multi-label classification tasks beyond food analysis.

Snakebite envenoming is a neglected tropical disease that affects millions of people each year, and species identification is a critical component in choosing the correct antivenom. In this study, six convolutional neural network and vision transformer architectures are compared for the image-based snake species classification task on the SnakeCLEF 2024 dataset, which contains 1599 snake species observed in 209 countries. We use knowledge distillation to boost accuracy without increasing inference cost, and incorporate country-level geographic location through a closed-form Bayesian fusion of the classifier’s output with location-conditioned species-occurrence priors. Location fusion reduced the classification error rate by over 40% across all architectures, and knowledge distillation provided a further improvement. The best model, CAFormer-S18, obtained an accuracy of 81.71% using both image and country-level location. We also compare the single-image inference latency on a smartphone and a Raspberry Pi 5, revealing that the latency of real-world, on-device deployments is significantly influenced by the interplay of model architecture and inference toolchain. Our top-performing model can infer in about 1 s on both devices, demonstrating that fine-grained species identification of snakes is feasible for real-time use in snakebite treatment and, when combined with an additional object detection stage, in wildlife monitoring.