Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 47 of the 48 references it cites were found in a registry or an index, and 1 are not checked yet; 4 of the 39 sources its claims cite were checked for support.

Deep Perceptual Hash Based on Hash Center for Image Copyright Protection

Open access
Sun, Xiaohan, Zhou, Jiting · 2022 · IEEE Access

Abstract

At present, most of the perceptual hash methods for image copyright protection rely on manually designed feature extraction and mapping, whose detection accuracy is insufficient. Some schemes based on deep learning are designed to consider a limited variety of content retention operations, which are not enough to deal with the increasingly severe situation of image copyright protection. In response to this situation, a novel Convolutional Neural Network (CNN)-based perceptual image hashing scheme is introduced in this paper. In this scheme, the training images are classified according to their original images and a Hadamard matrix is used to generate a hash center for each class in Hamming space. Then the Convolution Neural Network learns the feature extraction process of the image automatically, constrained by central quantization and distinct quantization, so that the hash code of each image converges to the hash center of their class, and generates the final hash sequence. The proposed scheme can successfully strike a balance between perceived robustness and discrimination capacity. Based on the test results on large-scale test sets,$F_{1}$scores, equal error rate (EER) and receiver operating characteristic (ROC) curves demonstrate the superiority of our scheme compared with some state-of-the-art schemes.

Verified 2026-10-10 · 1 reference not checked yet

Overview

CoChat Summary

TL;DR

The paper proposes a CNN-based perceptual image hashing scheme that treats transformed images from the same original as one class and uses pre-generated hash centers in Hamming space to learn compact hash codes for copyright protection. It reports better F1, EER, and ROC performance than several state-of-the-art schemes while requiring only 1,000 original training images.

Key findings

  • The proposed scheme achieved an F1 score of 0.9761 and EER of 0.0331 at 64 bits, compared with DSH (0.9463, 0.0609), CNMC (0.9450, 0.0772), IFNN (0.9101, 0.0894), RP-IVD (0.8367, 0.2014), and DBDH (0.8185, 0.1645).
  • It used only 1,000 original images for training and generated 84,000 transformed training images, whereas the paper states some prior methods require more than 200,000 original images.
  • On 800 different NUS-WIDE images, pairwise hash distances had a mean of 20.21 and standard deviation of 3.22, while transformed-image distances were mostly below 10 and only a few exceeded 14.06.
  • In stolen-image retrieval, when False Positive Ratio was 0.01, the scheme still detected more than 80% of stolen images; with 99,900 and 999,900 irrelevant images, recall decreased to 0.7256 and 0.6885, respectively.

Methods

  • The scheme classifies images by original source: images transformed from the same original image form one class, and a Hadamard matrix generates a hash center for each class in Hamming space.
  • ResNet50 extracts features, followed by a hash layer of three fully connected layers and ReLU; training is constrained by central quantization, distinct quantization, and a quantization loss.
  • Experiments used a newly built high-resolution copyright-protection dataset, USC-SIPI standard images, NUS-WIDE, and ImageNet; comparisons were made at 64 bits using F1, EER, ROC curves, execution time, and retrieval recall.

Limitations

  • The paper states that the quantization loss \(L_Q\) may have a negative effect in some cases and needs to be selected case-by-case; the ablation shows \(L_C\) is decisive and \(L_D\) significantly helps.
  • The paper states that recall decreases as the number of irrelevant images grows, with recall falling to 0.7256 and 0.6885 for 99,900 and 999,900 irrelevant images, indicating potential for further improvement.
  • The paper notes that Hadamard-matrix hash centers require the hash length to be a power of 2, so 24- and 48-bit hash sequences cannot be generated, though 16, 32, and 64 bits are sufficient.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/access.2022.3221980

Cites

47 of 48 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

FIGURE 1 Our pipeline: Images generated from the same original image are input as a same category. CNN is used for feature learning. Each category converges to its own hash center, after the processing of hash layers. Hash centers are pre-generated in Hamming space.
FIGURE 1 Our pipeline: Images generated from the same original image are input as a same category. CNN is used for feature learning. Each category converges to its own hash center, after the processing of hash layers. Hash centers are pre-generated in Hamming space.From the PDF
FIGURE 2 (a) Basic residual block adopted by ResNet18, 34; (b) Bottleneck residual block adopted by ResNet50, 101.
FIGURE 2 (a) Basic residual block adopted by ResNet18, 34; (b) Bottleneck residual block adopted by ResNet50, 101.From the PDF
FIGURE 3 Complete structure of ResNet50.
FIGURE 3 Complete structure of ResNet50.From the PDF
FIGURE 4 Some representative pictures in the dataset.
FIGURE 4 Some representative pictures in the dataset.From the PDF
FIGURE 5 Example of some transformed images.
FIGURE 5 Example of some transformed images.From the PDF
FIGURE 6 Five standard test images in the USC-SIPI dataset.
FIGURE 6 Five standard test images in the USC-SIPI dataset.From the PDF
FIGURE 7 Performance of perceptual robustness.
FIGURE 7 Performance of perceptual robustness.From the PDF
FIGURE 8 ROC curves of our scheme, RP-IVD [20], DSH [29], IFNN [33], CNMC [34] and DBDH [44].
FIGURE 8 ROC curves of our scheme, RP-IVD [20], DSH [29], IFNN [33], CNMC [34] and DBDH [44].From the PDF
FIGURE 9 ROC curves of our scheme when the hash code length is 16, 32, 64 and 128 bits.
FIGURE 9 ROC curves of our scheme when the hash code length is 16, 32, 64 and 128 bits.From the PDF
FIGURE 10 Recall Ratio curves for various schemes at different false positive ratios.
FIGURE 10 Recall Ratio curves for various schemes at different false positive ratios.From the PDF
Image 11
From the PDF
Image 12
From the PDF

Verification

Claims the paper makes

3 Valid · 2 Not valid · 16 Not checked · 56 Own results and interpretations

32 more claims

Create a free account to see all 77 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 30
  • Probably not a substantial claim. 16
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 45 references

Research works printed but not linked: 3

  1. weisstein, 2002, Hadamard matrix
  2. liu, 2020, An effective deep hash image copy detection algorithm, Applied Soft Computing, 3, 213
  3. 2016, The usc-sipi image database

Cited by

See all 20

Metadata

Authors
Sun, X., & Zhou, J.
Affiliations
Shanghai University (CN)
Venue
IEEE Access
Published
2022
Volume
10
Pages
120551-120562
ISSN
2169-3536 (ISSN-L)
OpenAlex
W4312538097
Cited by
20
References
45
Citations by year
2026: 3 · 2025: 5 · 2024: 5 · 2023: 7
FWCI
1.39
Type
journal article
Language
en
Open access
Locations
  • IEEE Access (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Steganography and Watermarking Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Retrieval and Classification Techniques
Keywords
perceptual image hashing, image copyright protection, Hadamard matrices, Hamming space, convolutional neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Deep Perceptual Hash Based on Hash Center for Image Copyright Protection. https://app.cochat.ai/papers/10.1109/access.2022.3221980
The paper
Sun, X., & Zhou, J. (2022). Deep Perceptual Hash Based on Hash Center for Image Copyright Protection. IEEE Access. https://doi.org/10.1109/access.2022.3221980

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.