Few-shot incremental food recognition via cross-domain guided pseudo-targets

The explosive growth of global food culture has expanded the application scope of visual recognition; however, it has introduced complex challenges arising from high intra-class variability and inter-class similarity. However, existing systems struggle to address fine-grained confusion and the trade-off between retaining old knowledge and adapting to new information. Traditional methods are constrained by a heavy reliance on large-scale datasets, whereas emerging zero-shot techniques are prone to semantic hallucination when encountering unseen dishes, thereby posing a severe challenge to precise recognition. To address these challenges, we propose the Cross-domain Guided Food Pseudo-Target Estimation (CFPE) framework, establishing a novel paradigm that is vision-led and semantically enhanced. First, to tackle the scarcity of incremental data, we utilize cross-domain adversarial training and an adaptive mask generator to synthesize high-quality pseudo-targets, thus establishing stable geometric anchors within the feature space. Second, by integrating Bessel Estimation Loss of Hypersphere (BELH) and Perturbation Margin Enhanced Prototype Regularization (PMEPR), we geometrically reconstruct the hyperspherical manifold distribution of features, effectively correcting estimation biases induced by few-shot samples. Crucially, we introduce a Food Factor-based Visual Semantic Consistency (FVSC) constraint, which explicitly decouples fine-grained visual confusion by injecting structured semantics. This is complemented by a depth-aware feature decoupling strategy to dynamically balance the plasticity and stability of the model. Experimental results demonstrate that CFPE achieves state-of-the-art performance across multiple benchmark datasets. It not only significantly improves incremental learning accuracy but also exhibits exceptional robustness in recognizing high-entropy food images.