Recognizing food ingredients in images is a challenging problem, particularly when models are intended for direct deployment on microcontrollers. Most related approaches formulate the problem as semantic or instance segmentation, which generally requires models with higher computational and memory demands, making such solutions often unsuitable for highly resource-constrained edge devices. This thesis therefore investigates the feasibility of using FOMO, a lightweight model, originally designed as a detection model for microcontrollers. We used the FoodSeg103 dataset, adapted the model to determine the presence of ingredients in an image, and evaluated the effects of dataset simplification, training settings, changes in input and output resolution, and quantization when preparing the model for deployment on the STM32 platform. Models from the YOLOv11-seg family were used as reference segmentation approaches, and the applicability of SAM methods was also assessed qualitatively. We found that FOMO is suitable in terms of model size, memory consumption, and feasibility of deployment on microcontrollers, but its performance remains limited for more demanding ingredient recognition tasks. The main limitations include a small receptive field, low-resolution output, limited model capacity, and sensitivity to data preparation and post-processing. The contribution of this thesis is an experimental assessment of the suitability of such an approach for microcontrollers and the derivation of guidelines for further development.
|
