Digital watermarking for camera-captured images based on just noticeable distortion and Wiener filtering

Highlights

• Securing the embedded watermark’s position and value is presented.

• A Gaussian based JND is proposed for adjusting the watermark strength.

• Extraction is achieved based on reducing distortions and predicting original image.

Abstract

In this paper, we propose a digital image watermarking method for camera-captured images. In our proposed method, an image component of all image pixels is used for embedding an individual watermark bit in order to provide large amount of the embedded watermark. The watermark strength is adjusted in accordance with the modified just noticeable distortion. After the watermarked image is printed and then captured by a digital camera, the reliable watermark extraction is accomplished based on the techniques of reducing distortions introduced from the printing and camera-capturing processes, and predicting original image component from the watermarked image component. In the experiments, various types of pixel value distortions and geometric distortions are considered and explored. With the proposed method, the results show that the watermark can be invisibly embedded, and reliably extracted. We also demonstrate its robustness against various types of distortions from the printing and camera-capturing processes.

Introduction

Digital media can be easily reproduced, especially digital images, and it is very difficult to distinguish illegal copies from the legal ones. This can lead to issues in many developing countries since it directly affects media owners/creators, and discourages them from inventing/developing future works. To address this problem, digital watermarking can be used to provide a mechanism for reliably verifying ownership of the distributed digital media. Digital image watermarking is used to embed secret digital information known as a “watermark” imperceptibly into a host digital image before distributing it to the public [1]. The quality degradation of a watermarked image must be unnoticeable by an observer of the image, while it is difficult for unauthorized person to modify, destroy or remove the embedded watermark. When the watermarked image is copied, the watermark will also be copied to the new image, and can later be extractable so the watermark can be “read” in order to provide electronic proof of ownership of such image [2]. This approach is commonly known as robust image watermarking, which is different from fragile image watermarking or image forensics approaches used for image authentication purposes [1].

The rapid growth of technology has led to the development of electronic devices. Particularly, printers and digital cameras are common devices used to publish and to reproduce documents. Digital images can be easily printed and rapidly distributed. For example, a digital image copied from Facebook and then printed on a poster or a printout contains such an image for an advertisement without permission from the owner (see Fig. 1). The copied and printed image is used and misunderstood by people that it belongs to someone else, not the real owner. Hence, his copyright, i.e. a legal right that grants the real owner of that image for its use, is violated. In this circumstance, if the original digital image contains a watermark, a digital version can be reproduced using a digital camera. In this paper we shall refer to this image as a “camera-captured image”, and investigate how an original image can be watermarked such that the embedded watermark will survive in a camera-captured version of the image. Note that the camera-captured image will be severely distorted from the printing and camera-capturing processes, resulting in an inaccuracy of the extracted watermark.

Various image watermarking methods have been proposed and shown to be robust against various types of distortions, e.g. rotation, scaling and translation (RST), shearing, etc. [3], [4], [5], [6], [7]. A few pioneer works on watermarking for camera-captured image has also been proposed. For example, a fast watermark detection method from a printed image proposed by Nakamura et al. [8] in 2004, where a camera phone was used to obtain a camera-captured image. With the help of a frame added around the image, the method was proved to be sufficiently robust against small geometric distortions. The captured image however needed to be aligned in front of the target image. Later, Takeuchi et al. [9] proposed a compensation method for lens and perspective-projective distortions in order to improve the correctness of watermark extracted from the captured image. Although the method was robust against both distortions, the effects of RST distortions were not considered and demonstrated. Kim et al. [10] proposed to embed a watermark repeatedly in form of a pseudorandom vector, while the amounts of the RST and perspective-projective distortions were reduced by using a tripod in the camera-capturing process, and by manually resizing and cropping the camera-captured image to its original size. Lee et al. [11] proposed a watermarking method based on local autocorrelation function (LACF) that was robust to Digital-to-Analog/Analog-to-Digital conversion (DAC/ADC), RST distortions, and perspective-projective distortion, which commonly occurred during the illegal copying of cinema footage. Conceptually, a watermark pattern was generated and added into part of the video frames in the spatial domain by taking an optimized Human Visual System (HVS). In detection process, the Weiner filter was used to predict the original video frame, while the LACF was used to estimate and recover geometric distortions for watermark synchronization purpose. Their method was experimentally proved to be robust against such distortions for video applications. In 2012, a periodic pattern-based watermarking method for interactive posters was proposed by Pramila et al. [12]. Basically, four watermark bits together with error control codes were used to generate particular 2-D patterns. The host image was divided into 9 successive blocks and used for patterns embedding separately. The technique based on enhanced peak detection with filtering was used to detect the angle of each embedded pattern, and the result was calculated and decoded to recover the four watermark bits. Although the method was proved to be robust against the RST and perspective-projective distortions, it suffered from a low embedding capacity, i.e. four watermark bits could only be embedded into each block of a host image.

In this paper, we present a new digital watermarking method for camera-captured images. Our method is designed to achieve both imperceptible watermark embedding and reliable watermark extraction from a printed image captured by a digital camera. In the watermark embedding process, a Gaussian based just noticeable distortion value, JND_G, is proposed for adjusting the watermark strength. In the watermark extraction process, three sub-processes are proposed, that is, an image registration technique based on a projective transformation is used for reducing geometric distortions, a color restoration technique based on the Color Mismatch Model with the RGB Speeded-Up Robust Features and local histogram equalization descriptors is used for decreasing pixel value distortions, and a prediction of original image component based on a modified 3 × 3 spatial domain Weiner filter is used for predicting original image component. Since various distortions are greatly introduced during the printing and camera-capturing processes, in the proposed watermarking method, the watermarked image is used to help decrease such distortions. In the next section, an overview of digital watermarking for camera-captured images is described. Details of our proposed watermark embedding and extraction processes are then presented in Section 3. Experimental settings are given in Section 4. The results are next shown and discussed in Section 5. The conclusion is in Section 6.