Convolutional Neural Network-Based Image Watermarking using Discrete Wavelet Transform Alireza Tavakoli1y Zahra Honjani2y and Hedieh Sajedi2

2025-04-26 0 0 4.04MB 10 页 10玖币

侵权投诉

Convolutional Neural Network-Based Image Watermarking

using Discrete Wavelet Transform

Alireza Tavakoli1†, Zahra Honjani2†, and Hedieh Sajedi2*

*Correspondence: hhsajedi@ut.ac.ir

1Department of Computer Engineering, School of Electrical and Computer Engineering, College of

Engineering, University of Tehran, Tehran

2Department of Computer Science, School of Mathematics, Statistics and Computer Science, College of

Science, University of Tehran, Tehran

†These authors contributed equally to this work.

Abstract

With the growing popularity of the Internet, digital images are used and transferred more

frequently. Although this phenomenon facilitates easy access to information, it also creates secu-

rity concerns and violates intellectual property rights by allowing illegal use, copying, and digital

content theft. Using watermarks in digital images is one of the most common ways to maintain

security. Watermarking is proving and declaring ownership of an image by adding a digital water-

mark to the original image. Watermarks can be either text or an image placed overtly or covertly

in an image and are expected to be challenging to remove. This paper proposes a combination

of convolutional neural networks (CNNs) and wavelet transforms to obtain a watermarking net-

work for embedding and extracting watermarks. The network is independent of the host image

resolution, can accept all kinds of watermarks, and has only 11 layers while keeping performance.

Performance is measured by two terms; the similarity between the extracted watermark and the

original one and the similarity between the host image and the watermarked one.

Keywords— watermarking, convolutional neural networks, wavelet transform, neural networks

1 Introduction

Protecting art ownership was always an issue since painters used to sign their works, which was not safe.

Due to widespread Internet use, artistic works are becoming digital, so the threat of privacy, copyright, and

illegal usage has increased. Authentication is not limited to artistic works but includes identity document (ID)

cards and source tracking. In order to solve these issues, digital watermarking is an eﬀective method of hiding

information that can be used for authentication purposes.

The digital watermarking technique involves embedding information (watermark) into the content and

extracting it when needed. Various attacks could be applied to the watermarked content to damage or remove

the embedded watermark, which shows the importance of robust watermarking. Even though watermarking

can be applied to a wide range of content, including video [1], audio [2], or images, in this paper, we will focus

on images.

Until recently, digital watermarking was based on deterministic algorithms. Typical methods are based

on embedding the watermark in one of the discrete cosine transform (DCT) [3], discrete wavelet transform

(DWT) [4–6], discrete Fourier transform (DFT) [7], or quantization index modulation [8], or a combination of

these [9,10].

Due to the inﬂexibility of deterministic algorithms, small changes in watermarked images lead to inaccurate

extracted watermarks. The inﬂexibility leads to using neural networks as a robust technique for digital water-

marking [11,12]. Even though these methods solve the robustness problem, some non-deterministic networks

still perform poorly, and others have complex structures.

This paper investigates the combination of deterministic and neural network methods to improve accuracy

and performance. As the results suggest, we have improvements in some attacks. In addition, the proposed

network enhances the quality of watermarked images by using DWT as a preprocessing technique.

The rest of the paper is organized as follows: In Section 2, we discuss relevant previous studies. Then in

Section 3, we outline the basic concept, wavelet transform, used for the proposed algorithm. The proposed

network structure is explained in Section 4. Section 5discusses the training technique, the experimental results,

and the comparison with literature [12], and [13]. Moreover, this paper is concluded in Section 6.

arXiv:2210.06179v2 [eess.IV] 29 Oct 2022

2 Related Works

Various image watermarking schemes have previously been done to provide robustness, authenticity, and im-

perceptibility. Earlier, we mentioned some papers proposing watermarking algorithms which use DWT, DFT,

and neural networks. Here we will explore some of them in more detail.

In paper [4], original images and watermarks are split into RGB channels to be used as inputs in the other

embedding process. In the original image, the blue channel has been subjected to DWT and Singular Value

Decomposition (SVD). A ﬁrst Arnold transformation is applied to the blue channel of the watermark image,

which acts as a key. In addition to Arnold’s transformation, SVD is applied to the watermark image, and then

the image is embedded with the watermark. In addition to being semi-blind, this method is also robust against

attacks due to the use of the key during the watermark extraction.

As another deterministic approach, Xiangui Kang proposed a blind DWT-DFT composite image water-

marking algorithm that is robust against both aﬃne transformation and Joint Photographic Experts Group

(JPEG) compression [10]. This paper used watermark structure, two-dimensional interleaving, and synchro-

nization techniques to improve the robustness.

In recent years, residual blocks, fully connected layers, DCT layers, and CNN have been used as watermark-

ing networks. An end-to-end autoencoder in ReDMark [12] simulates a DCT layer in its deep network. The

concatenation of the input image and watermark is used as input to the DCT layer, and after transformation,

the embedding is performed. The proposed structure is robust against JPEG attacks since the DCT layer is

used.

Jae-Eun Lee proposed a robust watermarking neural network that included an attack simulation and was

adjusted to the resolution of the host image and the watermark [13]. It is composed of convolutional layers with

average pooling layers. This method also includes a strength factor to adjust the tradeoﬀ between invisibility

and robustness. This method uses the images’ luma component (Y) to hide the watermark. A watermark is

converted into a layer the same size as the original image. The watermark is then merged with the processed

image in the leading network. Despite the impressive results, the proposed network is relatively complex. In

contrast to mentioned articles, this paper presented a non-deterministic method with higher robustness than

discussed papers.

3 Discrete Wavelet Transform

Each image has smooth regions interrupted by edges. Edges provide much information about the image.

In the Fourier transform, a signal is decomposed into sin waves that are not localized in time. Therefore,

Fourier transforms cannot eﬃciently represent sudden changes (particularly edges), and wavelet transforms

are preferred. A wavelet transform is a way of representing a function by a speciﬁc orthonormal series. A

wavelet is a wave-like oscillation that is localized in time. Diﬀerent types of wavelet transform exist based on

the mother function ψ(t) and the scaling function ϕ(t). One of the most widely-used wavelets is Haar whose

ψ(t) and ϕ(t) are shown in Figure 1.

(a) Scaling function (b) Mother function

Figure 1: Haar’s functions.

In addition, the discrete wavelet transform decomposes an image into four wavelet subbands, LL, LH, HL,

and HH, which were examined independently to compare their capacity for hiding information. The results

are demonstrated in Section 5[14].

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ConvolutionalNeuralNetwork-BasedImageWatermarkingusingDiscreteWaveletTransformAlirezaTavakoli1y,ZahraHonjani2y,andHediehSajedi2**Correspondence:hhsajedi@ut.ac.ir1DepartmentofComputerEngineering,SchoolofElectricalandComputerEngineering,CollegeofEngineering,UniversityofTehran,Tehran2DepartmentofComput...

展开>> 收起<<

Convolutional Neural Network-Based Image Watermarking using Discrete Wavelet Transform Alireza Tavakoli1y Zahra Honjani2y and Hedieh Sajedi2.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Convolutional Neural Network-Based Image Watermarking using Discrete Wavelet Transform Alireza Tavakoli1y Zahra Honjani2y and Hedieh Sajedi2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: