RFHNet: Relational and Frequency-Aware Hashing Network for Large-Scale Fine-Grained Food Image Retrieval

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Food Computing · Depth: Expert, extended

Summary

RFHNet, a cascaded hierarchical hashing network, is proposed for large-scale fine-grained food image retrieval, a critical task in computational gastronomy. This network addresses limitations of existing methods by capturing subtle local semantics and frequency-sensitive visual cues through multi-level representations. RFHNet integrates three core components: Fine-grained Relation Modeling (FRM) for subtle visual differences, Multi-Frequency Modulated Fusion (MFMF) for multi-frequency feature extraction, and Hierarchical Semantic Synergy (HSS) for adaptive multi-level representation integration. Evaluated on six food-specific benchmarks including Food-101 and Food2K, RFHNet consistently outperforms state-of-the-art hashing methods, achieving mAP gains ranging from 4.44% to 17.20% at 12 bits. It also exhibits superior inference efficiency compared to Transformer-based FoodHash.

Key takeaway

For Machine Learning Engineers developing large-scale fine-grained image retrieval systems, particularly for food computing, you should consider RFHNet. Its ability to model subtle spatial relations and multi-frequency features, combined with hierarchical semantic synergy, significantly improves retrieval accuracy. This approach offers superior mAP gains and better inference efficiency than prior methods. It is a strong candidate for your next-generation smart catering or dietary monitoring applications.

Key insights

RFHNet enhances fine-grained food image retrieval by integrating relational and multi-frequency modeling with hierarchical semantic fusion for discriminative hash codes.

Principles

• Capture subtle spatial relations between local regions.
• Exploit multi-frequency features to suppress noise.
• Adaptively integrate hierarchical features for robust fusion.

Method

RFHNet uses a cascaded hierarchical network with a ResNet-50 backbone. It applies Fine-Grained Relational Modeling (FRM) and Multi-Frequency Modulated Fusion (MFMF) to intermediate features, then Hierarchical Semantic Synergy (HSS) for adaptive cross-scale fusion, generating hash codes.

In practice

• Apply cascaded hashing for fine-grained image tasks.
• Use 2D-FFT and dynamic gating for frequency features.
• Integrate relational modeling for subtle visual cues.

Topics

• Fine-Grained Image Retrieval
• Hashing Networks
• Food Computing
• Multi-Frequency Features
• Relational Modeling
• Deep Hashing

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• RFHNet: Relational and Frequency-Aware Hashing Network for Large-Scale Fine-Grained Food Image Retrieval — RFHNet improves fine-grained food image retrieval by combining relational modeling, multi-frequency feature extraction, and hierarchical semantic synergy.
• Batch-Invariant Spectral Intelligence for Robust and Explainable Insect Authentication — BISN uses learnable preprocessing and adversarial training to achieve robust, explainable insect authentication despite batch variations.
• TACoS: Weakly Supervised Learning of Two-Dimensional Materials from Scribble Annotations to Precise Segmentation — TACoS enables precise 2D material segmentation using minimal scribble annotations via a novel semi-supervised framework.
• Frequency-Enhanced Dual-Subspace Networks for Few-Shot Fine-Grained Image Classification — FEDSNet improves few-shot fine-grained image classification by integrating frequency domain features to enhance structural stability and reduce overfitting.
• Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces — Reasoning in large output spaces can be mechanistically characterized as a two-phase process: shortlisting followed by fine-grained reasoning.
• Smart Data Grouping: Organizing Networks Without Guesswork — A new differentiable structural information formulation enhances deep graph clustering via LSEnet.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.