Source-linked AI summary
Deep Supervised Discrete Hashing
Qi Li, Zhenan Sun, Ran He, Tieniu Tan
TL;DR
Previous deep hashing methods may not fully exploit semantic information because classification learning does not directly affect the hash function. This paper jointly learns directly binary hash codes from pairwise and classification information in one stream, outperforming state-of-the-art methods on benchmark image-retrieval datasets.
Problem
Previous deep hashing methods may not fully exploit semantic information, while two-stream classification learning does not directly affect the hash function.
Method
The method jointly learns directly binary hash codes using pairwise label and classification information in one CNN stream, optimized with alternating minimization.
Results
Across benchmark image-retrieval datasets, the method outperforms state-of-the-art methods, reaching an average MAP of 0.938 on CIFAR-10 under the second experimental setting.
Takeaways & Limitations
Jointly learning the hash function and classifier within one stream can boost retrieval performance.
Takeaways & Limitations
The approach assumes that learned binary codes should be ideal for classification.
Abstract
from arXiv · showhide
With the rapid growth of image and video data on the web, hashing has been extensively studied for image or video search in recent years. Benefit from recent advances in deep learning, deep hashing methods have achieved promising results for image retrieval. However, there are some limitations of previous deep hashing methods (e.g., the semantic information is not fully exploited). In this paper, we develop a deep supervised discrete hashing algorithm based on the assumption that the learned binary codes should be ideal for classification. Both the pairwise label information and the classification information are used to learn the hash codes within one stream framework. We constrain the outputs of the last layer to be binary codes directly, which is rarely investigated in deep hashing algorithm. Because of the discrete nature of hash codes, an alternating minimization method is used to optimize the objective function. Experimental results have shown that our method outperforms current state-of-the-art methods on benchmark datasets.
1 Introduction
Deep hashing methods have advanced image retrieval but may not fully exploit semantic information. This work addresses that limitation by learning directly binary hash codes that jointly preserve similarity relationships and label consistency within one stream.
- Motivation: Hashing encodes high-dimensional images or videos into binary codes while preserving similarity, enabling computationally inexpensive and storage-efficient retrieval.The introduction motivates hashing through the rapid growth of image and video data on the web.
- Limitations: Previous deep hashing methods can leave semantic information underexploited, while two-stream approaches separate hash-function learning from semantic mining.The two streams are described as a hash stream and a classification stream under a multi-task learning framework.
- Proposed method: The method constrains its last layer to output binary codes directly, while simultaneously preserving similarity relationships and maintaining label consistency.The contribution states that this direct binary output is intended to avoid learning continuous outputs before binarization.
- Proposed method: The work jointly uses pairwise label information and classification information to learn hash codes under one stream framework.The authors identify this combination as, to their knowledge, the first of its kind among deep hashing methods.
2 Deep supervised discrete hashing
Deep supervised discrete hashing jointly learns binary codes from pairwise similarity and classification information, under the assumption that codes should be ideal for classification. It directly constrains the network’s final-layer outputs to binary codes and optimizes the resulting discrete objective by alternating minimization.
- Model formulation: For similar image pairs, the loss minimizes Hamming distance, whereas for dissimilar pairs, it maximizes Hamming distance.The pairwise likelihood uses binary-code inner products as a similarity measure.
- Model formulation: The method combines pairwise label information with a jointly learned linear classifier so binary codes preserve similarity and support classification.Its objective maximizes pairwise label likelihood while optimizing the learned classifier jointly.
- Discrete optimization: The network constrains outputs of the last fully connected layer to be binary codes directly, avoiding relaxation through continuous outputs and thresholding.The paper argues that ignoring binary constraints can produce suboptimal codes because Euclidean and Hamming spaces differ.
- Discrete optimization: An auxiliary variable reformulates the discrete objective into two subproblems solved iteratively by alternating minimization.A sufficiently large Lagrange multiplier penalizes constraint violations and forces final-layer outputs closer to the binary codes used for classification.
- Discrete optimization: The discrete cyclic coordinate descent procedure updates the hash-code matrix row by row until convergence, with each bit computed from the other K −1 bits.The classifier parameters and network parameters are updated in alternating optimization steps.
3 Experiments
Experiments on CIFAR-10 and NUS-WIDE evaluate DSDH against traditional and deep hashing methods using multiple retrieval metrics and controlled variants. DSDH benefits from incorporating classification information and achieves strong MAP performance, especially on CIFAR-10.
- Datasets and evaluation: Experiments use CIFAR-10 and NUS-WIDE, including single-label and multi-label image retrieval benchmarks.CIFAR-10 contains 60,000 images in 10 classes, while NUS-WIDE contains 269,648 images with 5,018 tags; a 195,834-image subset is used for NUS-WIDE.
- Datasets and evaluation: Retrieval quality is evaluated with MAP, Hamming-radius precision, top-return precision, and precision-recall curves.For NUS-WIDE MAP, the evaluation considers the top 5,000 returned neighbors in the first setting and top 50,000 in the second.
- Ablation studies: DSDH-C outperforms DSDH-A, while DSDH-B also improves over DSDH-A on Hamming-radius-2 precision and precision-recall curves.These variants test the effects of classification information, a two-stream design, and directly applying the sign function to the final-layer outputs.
- Comparison with baselines: On CIFAR-10, DSDH substantially outperforms traditional hashing methods, with MAP more than twice that of SDH, FastH, and ITQ.Most deep hashing methods also outperform traditional methods, although DTSH achieves the best performance among the other methods except the proposed method.
- Comparison with baselines: On NUS-WIDE, SDH averages 0.603 MAP and DTSH averages 0.804, while DSDH is slightly superior to DTSH.The deep-versus-traditional gap is smaller on NUS-WIDE, which has more categories and multiple labels per image.
4 Conclusion
The paper proposes a deep supervised discrete hashing algorithm that directly learns binary codes using pairwise label and classification information in one stream, optimized by alternating minimization.
- 4 Conclusion: The method directly constrains the last-layer outputs to binary codes while jointly using pairwise label and classification information.These components are integrated within a one-stream framework for learning hash codes.
- 4 Conclusion: An alternating minimization method optimizes the loss function to address the discrete nature of the hash codes.
- 4 Conclusion: Extensive experiments show that the proposed method outperforms state-of-the-art methods.