Progressive Generative Hashing for Image Retrieval

Yuqing Ma, Yue He, Fan Ding, Sheng Hu, Jun Li, Xianglong LiuIJCAI 2018 · DOI

Abstract

Efficient large-scale image retrieval relies on compact binary hash codes that preserve semantic similarity. While supervised hashing methods excel by leveraging labels, unsupervised hashing remains challenging due to the difficulty of capturing high-level semantics from raw pixels. We present Progressive Generative Hashing (PGH), an unsupervised framework that integrates Generative Adversarial Networks (GANs) to discover latent semantic structures. By conditioning image generation on binary codes and iteratively refining the hashing network through a progressive learning loop, PGH "hallucinates" semantic anchors that guide the learning process. Experimental results on MNIST and CIFAR-10 demonstrate that PGH significantly outperforms existing unsupervised deep hashing methods.

Introduction

In the era of massive visual datasets, hashing has become indispensable for rapid image retrieval. The goal is to map high-dimensional images into short binary strings such that semantically similar images have similar codes. While supervised methods benefit from human-annotated labels, they are often impractical for large-scale, unlabeled web data. Unsupervised hashing methods typically rely on data distribution or reconstruction losses, which often fail to capture complex semantic relationships.

Progressive Generative Hashing (PGH) addresses this by using a GAN-based architecture to bridge the gap between low-level pixels and high-level semantics. The core idea is to treat hash codes as semantic conditions for image generation, allowing the model to learn a mapping that is consistent across both real and synthetic domains.

Methodology

The PGH framework consists of three primary components: a Hashing Network ($H$), a Generator ($G$), and a Discriminator ($D$).

The PGH architecture illustrating the interaction between the Hashing Network (H), Generator (G), and Discriminator (D) through a progressive learning loop.

Hash-Conditioned GANs

Unlike standard GANs that generate images from noise, the PGH generator $G$ is conditioned on a binary hash code $B^{(t-1)}$ and a noise vector $z$. This forces the generator to associate specific binary patterns with visual semantic features. The discriminator $D$ ensures that the generated synthetic images $\bar{X}$ are indistinguishable from real images $X$.

Progressive Learning Loop

The "Progressive" nature of PGH stems from its iterative training process. The model starts with initial codes generated by a simple baseline method. These codes are used to train the GAN. Once the GAN can generate semantically relevant synthetic images, these images are used to refine the hashing network $H$ via a triplet loss. This loss encourages the network to map a real image $X$ and its corresponding synthetic version $\bar{X}$ to similar codes, while pushing dissimilar synthetic images away. The updated hashing network then produces better codes $B^{(t)}$, which are fed back into the GAN for the next cycle of refinement.

Experimental Results

The authors evaluated PGH on the MNIST and CIFAR-10 datasets, comparing it against a variety of shallow (e.g., LSH, ITQ) and deep unsupervised hashing methods (e.g., Deepbit, BGAN).

Performance comparison (PH2) of PGH against various baselines on MNIST and CIFAR-10 datasets across different bit lengths.

Quantitative Performance

As shown in the performance charts, PGH consistently achieves superior results. On MNIST, PGH reaches a mean Average Precision (mAP) of 67.95% at 64 bits, significantly higher than BGAN (51.61%). On the more complex CIFAR-10 dataset, PGH demonstrates a substantial lead, achieving 33.40% mAP at 16 bits, outperforming Deepbit (19.43%) and ITQ+VGG (25.51%).

Semantic Discovery

The progressive refinement allows PGH to discover latent categories. Visualizations of the learned hash space indicate that PGH naturally clusters semantically related classes (e.g., different types of vehicles in CIFAR-10) more effectively than competing unsupervised methods. This suggests that the GAN-driven "hallucination" process successfully acts as an automated labeler, providing the necessary guidance for semantic hashing.

Conclusion

Progressive Generative Hashing offers a novel approach to unsupervised image retrieval by leveraging the generative power of GANs. By iteratively bootstrapping the hashing network and the generator, the model learns compact codes that capture semantic similarity without human intervention. While the computational cost of training GANs progressively is a consideration, the significant gains in retrieval accuracy make PGH a compelling framework for large-scale unsupervised learning.

Summary updated

BibTeX
@inproceedings{ma2018progressive,
  author = {Yuqing Ma and Yue He and Fan Ding and Sheng Hu and Jun Li and Xianglong Liu},
  title = {{Progressive Generative Hashing for Image Retrieval}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/121},
}
Figure