Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis

Yinan Li, Jun Long, Zhan YangAAAI 2025 · DOI

Introduction

Cross-modal retrieval—the ability to search for images using text queries or vice versa—is a cornerstone of modern big data systems. To achieve real-time performance, hashing techniques are employed to map high-dimensional data into compact binary codes. However, a persistent "semantic gap" remains: traditional hashing often fails to capture the hierarchical and nuanced relationships between different semantic concepts, leading to coarse retrieval results.

The ACHFCA Framework

Researchers from the Big Data Institute at Central South University have proposed a novel framework: Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis (ACHFCA). This approach leverages mathematical lattice theory to bridge the semantic gap more effectively.

The architecture is built around a "flash-projection" semantic enhancement descriptor. As illustrated in Figure 1, the framework consists of three primary layers:

  1. Semantic Hash Mapping Layer: Maps semantic labels into primary binary codes.
  2. Feature Alignment Layer: Aligns latent representations of heterogeneous data (images and text) to mitigate noise.
  3. Hash Code Generation Layer: Synthesizes variables from the previous layers to produce final hash codes.
The ACHFCA framework architecture illustrating the three-layer descriptor and the concept lattice integration.

Fine-Grained Similarity via Concept Lattices

The core innovation of ACHFCA is the use of Formal Concept Analysis (FCA). By treating labels as a "formal context," the model constructs a concept lattice. Unlike traditional methods that treat labels as binary (similar or not), the lattice allows for the calculation of fine-grained similarity based on structural positions, such as nearest common parents and shortest paths. This "humanized" approach captures the inherent hierarchy of concepts (e.g., "sunset" relating to "sky").

Asymmetric Learning and Optimization

To preserve modality-specific information, ACHFCA utilizes an asymmetric learning framework. Instead of forcing images and text into an identical space prematurely, real-valued representations are used to approximate binary codes. A discrete optimization algorithm—specifically an alternating direction minimization approach—is used to learn hash codes directly, avoiding the precision loss associated with continuous relaxation.

Experimental Performance

The framework was evaluated on three major benchmarks: MIRFlickr, NUS-WIDE, and IAPR-TC12. ACHFCA consistently outperformed state-of-the-art (SOTA) methods such as BATCH, SRLCH, and ASCSH across various code lengths.

Precision-Recall curves for the MIRFlickr dataset (Text-to-Image task) at 16 bits, showing ACHFCA's superior performance.

On the MIRFlickr dataset, ACHFCA achieved a Mean Average Precision (mAP) of 0.7831 at 64 bits for Image-to-Text tasks. Across all datasets, the model showed an average mAP improvement of 1.3% to 1.9% over previous supervised SOTA methods. Furthermore, the optimization algorithm demonstrates high efficiency, typically converging within 10 to 50 iterations.

Conclusion and Future Directions

ACHFCA demonstrates that incorporating hierarchical semantic structures via Formal Concept Analysis significantly enhances cross-modal retrieval. While performance remains strong, the authors noted that very high-dimensional deep features can present challenges for semantic projection. Future work will explore three-way concept analysis and self-supervised learning to further refine the similarity matrix and reduce reliance on manual annotations.

Summary updated

BibTeX
@inproceedings{li2025asymmetric,
  author = {Yinan Li and Jun Long and Zhan Yang},
  title = {{Asymmetric Cross-Modal Hashing Based on Formal Concept Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i2.32129},
}
Figure