Mask-guided multistage food recognition via adaptive segmentation and efficient classification

Recent advances in computer vision have enabled fine-grained food understanding, which plays a vital role in dietary monitoring, nutrition estimation, and smart catering applications. However, food recognition remains challenging due to strong intra-class variations, complex backgrounds, and frequent occlusions. This paper proposes a multi-stage food recognition framework that combines semantic segmentation and classification in a unified pipeline. A U-Net network is first employed to generate precise food masks, which guide the subsequent classification stage based on EfficientNet-B3. A mask-guided cropping mechanism is introduced to suppress background noise and focus the classifier on relevant regions. Comprehensive experiments on the UEC-FoodPix Complete dataset demonstrate that the proposed framework achieves F1-macro of 0.706 ± 0.009, outperforming single-stage baselines (EfficientNet-B3: 0.672 ± 0.012, Swin-B: 0.685 ± 0.010) and detection-based methods (Faster R-CNN + EfficientNet: 0.689 ±0.011, Mask R-CNN: 0.698 ±0.009). Cross-dataset evaluation on Food-101 validates the framework's generalization capability with only 7.6% performance drop compared to 10-12% for baseline methods. These results highlight the effectiveness of integrating pixel-level segmentation and image-level classification for robust food recognition, paving the way for future multi-modal food analysis and lightweight deployment.