Аннотация
Automated food detection is a major component in building reliable dietary monitoring systems, particularly in regions where cuisine-specific datasets are limited. Central Asian food recognition presents unique challenges due to strong intra-class variation, visually similar dishes, multi-item meal compositions, and long-tailed class distributions. To address this problem, we present a comprehensive benchmarking analysis of five modern object detection models across three groups: two-stage (Faster R-CNN), one-stage anchor-based (RetinaNet), one-stage anchor-free (YOLOv12, YOLO26), and transformer-based (RT-DETR). Experiments are conducted on a unified dataset constructed by merging the Global Gastronomic Culinary Dataset (GGCD) and the Food Portion Benchmark (FPB), resulting in 48,228 images across 258 food classes. The dataset includes controlled single-portion photographs and multi-item meals captured across various backgrounds and from different viewing angles. The merged dataset therefore covers both isolated food portions and more complex scenes containing several dishes. We compare the models in terms of detection accuracy, computational requirements, and inference speed, and examine AP by food class across all models. Overall, YOLOv12 achieves the highest accuracy, with an mAP50 of 75.80 and an mAP50-95 of 67.90, while using fewer parameters and less computation than the heavier baselines. RT-DETR reaches an mAP50 of 72.90 but requires more computation, whereas YOLO26 records the shortest single-image inference time on both GPU and CPU. Class-wise analysis further shows high performance on visually distinctive categories, with near-perfect AP for ten top-detected classes. These findings provide a practical baseline for Central Asian food detection and show that future progress depends on improving robustness for long-tail and visually ambiguous classes.