In university environments, canteens serve thousands of students daily, especially during short breaks when many must obtain meals in limited time. At the King Fahd University of Petroleum and Minerals (KFUPM) students’ canteen, automating the food purchasing process using computer vision can enhance the student experience and save considerable time and effort. Although many food datasets exist, most were not collected under real canteen conditions, where multiple food classes appear on the same tray and often overlap. Moreover, existing tray-based datasets include food classes different from those served at the KFUPM students’ restaurant, limiting their direct applicability. This research introduces KF-2025, the first real-world tray-based food image dataset collected directly from the KFUPM students’ canteen, designed to support automated self-checkout and smart cafeteria applications. The dataset was recorded as videos and decomposed into frames, capturing over 1,000 trays. A total of 711 images were manually annotated using Roboflow and cropped to produce 3,222 food item images across 26 classes. Four transfer learning models (ResNet-50, ResNet-152, MobileNetV2, and EfficientNet-B0) were trained and evaluated, with ResNet-152 achieving the best performance using the Adam optimizer: 97.25% accuracy, 97.62% precision, 93.78% recall, and 94.53% F1-score. The results confirm KF-2025’s reliability and real-world potential, as even lightweight models exceeded 94% accuracy, demonstrating feasibility for deployment on resource-constrained devices such as Jetson or Raspberry Pi and establishing KF-2025 as a solid foundation for future smart canteen applications.