FOCUS: A Frequency-Oriented and Class-Underrepresented Semantic Segmentation Framework for Food Images

ICLR 2026 Conference Submission19205 Authors

Keywords: Food computing, Semantic segmentation, Food segmentation, Food recognition

Abstract: Generating high-quality semantic segmentation results from food images remains a challenging task, particularly in the presence of complex boundaries and class imbalance. Existing methods often struggle with blurred edges and underperform on long-tailed categories, limiting their generalizability in practical scenarios. To address these issues, we propose FOCUS, a novel semantic segmentation framework designed to enhance boundary precision and improve rare class recognition. Specifically, we introduce a frequency-based strategy that selectively processes high-frequency components via differential convolution and integrates explicit edge supervision during training. This enables the model to better capture fine-grained boundary details and improves edge discriminability. To mitigate class imbalance, we introduce an enhanced gradient allocation mechanism that applies targeted matching supervision to rare categories, thereby amplifying learning signals for low-shot classes and improving classification accuracy. Extensive experiments on benchmark datasets, FoodSeg103, UECFoodComplete, and Food50Seg, demonstrate that FOCUS consistently outperforms existing approaches in both boundary quality and rare class performance, validating its architectural effectiveness and robust generalization capability.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19205