Research map / Vision-language and multimodal / Cross-modal learning
Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ACHFCA, an asymmetric supervised cross-modal hashing framework that builds a three-layer semantic descriptor and a fine-grained similarity matrix from formal concept analysis (concept lattice) with discrete optimization.
Problem
Cross-modal hashing methods fail to fully exploit semantic labels to build association relations between sample pairs, limiting retrieval accuracy on multimodal image-text retrieval.
Approach
Uses formal concept analysis over the label matrix to compute a fine-grained pairwise similarity (combining shared attributes and concept-lattice position), and learns asymmetric hash codes from different layers of a three-layer linear semantic descriptor via alternating discrete optimization.
Key results
- ACHFCA improves average I2T mAP over the best baseline by 1.68% (MIRFlickr), 1.77% (NUS-WIDE) and 1.83% (IAPR-TC12).
- For T2I, average mAP is higher by 1.62%, 1.35% and 1.89% on the three datasets respectively.
- With a deep-feature variant (ACHFCA-D, VGG features), T2I mAP exceeds DCHMT and DSPH, but I2T mAP is lower than those end-to-end deep baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · MIRFlickr I→T, 16 bits | mAP: 0.7508 -> 0.7605 | Helped |
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · MIRFlickr T→I, 16 bits | mAP: 0.8133 -> 0.8225 | Helped |
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · NUS-WIDE I→T, 16 bits | mAP: 0.6413 -> 0.6492 | Helped |
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · NUS-WIDE T→I, 16 bits | mAP: 0.7565 -> 0.7618 | Helped |
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · IAPR-TC12 I→T, 16 bits | mAP: 0.5110 -> 0.5176 | Helped |
| Use ACHFCA instead of the strongest baseline | Cross-modal hashing · IAPR-TC12 T→I, 16 bits | mAP: 0.5637 -> 0.5715 | Helped |
| Add RBF-kernelized modality features (ablation) | ACHFCA · MIRFlickr I→T | mAP: 0.7605 -> 0.7278 | Hurt |
| Add RBF-kernelized modality features (ablation) | ACHFCA · MIRFlickr T→I | mAP: 0.8225 -> 0.7324 | Hurt |
| Add RBF-kernelized modality features (ablation) | ACHFCA · NUS-WIDE I→T | mAP: 0.6492 -> 0.6290 | Hurt |
| Add RBF-kernelized modality features (ablation) | ACHFCA · NUS-WIDE T→I | mAP: 0.7618 -> 0.6647 | Hurt |
| Replace FCA similarity reconstruction with classical label similarity (ablation) | ACHFCA · MIRFlickr I→T | mAP: 0.7605 -> 0.7418 | Hurt |
| Replace FCA similarity reconstruction with classical label similarity (ablation) | ACHFCA · MIRFlickr T→I | mAP: 0.8225 -> 0.7937 | Hurt |
Surprising findings
- The RBF kernelized variant (ACHFCA-K) causes a large performance drop, especially for T2I.
- With VGG deep image features, ACHFCA-D underperforms end-to-end deep baselines on I2T tasks.
Limitations
- Relies on linear hash mappings; the deep-feature variant underperforms end-to-end deep baselines on I2T.
- Kernelization hurts performance, indicating the approach is not robust to nonlinear feature mapping.
- Evaluated only on three datasets with a limited set of baselines and hash lengths.
Experimental setup
Datasets: MIRFlickr NUS-WIDE IAPR-TC12
Benchmarks: MIRFlickr NUS-WIDE IAPR-TC12
Metrics: mAP precision-recall curves
Baselines: BATCH SRLCH ASCSH ALECH HCCH ROHLSE DCHMT DSPH
Builds on: asymmetric semantic hashing formal concept analysis three-way concept analysis
Closest papers by meaning
- Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (AAAI 2025 · similarity 0.90)
- Dynamic Masking and Auxiliary Hash Learning for Enhanced Cross-Modal Retrieval (NeurIPS 2025 · similarity 0.90)
- Stationary and Clustering Transformer Hashing for Cross-modal Retrieval (AAAI 2026 · similarity 0.90)
- Statistical Model-driven Similarity Hashing: Bridging Modalities for Efficient Unsupervised Retrieval (AAAI 2025 · similarity 0.88)
- Vision-guided Text Mining for Unsupervised Cross-modal Hashing with Community Similarity Quantization (AAAI 2025 · similarity 0.88)
- Robust Self-Paced Hashing for Cross-Modal Retrieval with Noisy Labels (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.