Adaptive self-recovery for tampered images based on VQ indexing and inpainting

Abstract

In this paper, we propose a novel self-recovery scheme for tampered images using vector quantization (VQ) indexing and image inpainting. Cover image blocks are classified into complex blocks and smooth blocks according to the distribution characteristics. Due to the good performance of the compressed representation of VQ and the automatic repairing capability of image inpainting, the recovery-bits of each cover block are generated by its VQ index and the inpainting indicator. Recovery-bits and authentication-bits are embedded into the LSB planes of the cover image to produce the watermarked image. On the receiver side, after tampered blocks are all localized, the extracted recovery-bits are used to judge the classification of each tampered block. By analyzing the validity of the VQ indices and the damaged degree of the neighboring regions, the adaptive recovery mechanism can be utilized to restore all the tampered blocks by using VQ index and image inpainting. Experimental results demonstrate the effectiveness of the proposed scheme.

Introduction

With the rapid development of the Internet and digital signal processing technologies, various processes, such as editing, copying, and distribution of digital contents, have become more and more convenient. However, if the digital content is distributed illegally or tampered maliciously, copyright infringements and harmful social effects could occur. Therefore, how to protect digital contents has aroused the significant interest among the academia and the industry [1], [2], [3], [4], [5], [6], [7].

Currently, there are many techniques that can be used to identify the trustworthiness and the integrity of digital content. The multimedia hashing technique can generate a fixed-length sequence that is a compressed representation of the principle features of the multimedia data [8], [9], [10]. Multimedia data that are perceptually similar produce similar hash sequences, whereas perceptually different data have very distinct hash sequences. Thus, multimedia hashing can be used for authentication. However, the hash sequence must be appended with the original multimedia data and transmitted to the receiver together in order to compare it with the hash of the received multimedia data. The digital forensic technique can be used to determine whether received multimedia data have undergone certain malicious manipulations without having any information about the original data [11], [12], [13]. Intrinsic traces and inconsistencies, such as the color filter array (CFA) of image capturing devices and lighting directions, are analyzed to produce the authenticity judgment. However, such forensic schemes have relatively low accuracy and involve considerable computational complexity. The fragile watermarking technique can achieve multimedia authentication by embedding the auxiliary information imperceptibly, i.e., watermark, into the multimedia cover data [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27]. The integrity of the received data can be judged easily according to the extracted watermark. If the embedded watermark is generated from the multimedia cover data, such schemes are called self-embedding fragile watermarking schemes. In this work, our main focus is on fragile watermarking of digital images.

The aim of the earlier research on the fragile image watermarking was to realize the localization of the modified or tampered image regions [14], [15], [16], [17], [18]. There are two categories of fragile watermarking, i.e., block-wise schemes and pixel-wise schemes, which differ in their accuracy of locating the modified or tampered regions.

Block-wise schemes often divide the cover image into non-overlapping blocks and embed the watermark into each block [14], [15], [16], and the embedded watermark can be a hash of the principal content of each cover image block. If the watermarked image is tampered by an adversary, the extracted watermark and the image content corresponding to the tampered blocks are not matched so that the localization of the tampered blocks can be achieved. Pixel-wise schemes often generate the watermarked image by embedding the watermark information derived from the cover image pixels [17], [18]. A pixel-wise fragile watermarking based on a statistical mechanism was proposed in [18]. In this method, a set of tailor-made authentication data for each cover pixel and some additional test data were embedded into the cover image. On the authentication side, two different distributions of tampered and original pixels can be used to locate the tampered pixels. The pixel-wise scheme has more precise localization capability than the block-wise scheme. However, its limitation is that it allows only a relatively small tampered region.

Recent research on fragile watermarking methods has emphasized the recovery of image content in addition to the tampering localization [19], [20], [21], [22], [23], [24], [25], [26], [27]. These reported methods often embed a compact representation of the image content into the cover image. Once the watermarked image is tampered, the extracted watermark can be decoded and used for content recovery.

Fridrich et al. [19] encoded the DCT coefficients of each cover image block into 64 or 128 bits and used them to replace the least significant bits (LSB) of another block. After the tampered blocks were identified, the quantized DCT coefficients were extracted from the reserved regions and decoded to recover the principal content of the tampered areas. Zhang et al. [20] incorporated the block-wise scheme and the pixel-wise scheme by using a hierarchical mechanism. After identifying the tampered blocks, the watermark bits embedded in the intact blocks were exploited to locate the tampered pixels. But the recovery procedure was based on exhaustive attempts, which are time-consuming and unrealistic in real-world applications. In [21], by a reversible data hiding technique, reference-bits and check-bits were embedded into the cover image as watermark information. When the tampered region is not too large, this scheme can recover the original image with no errors, but the visual quality of the watermarked image is unsatisfactory. A fragile watermarking scheme that has the capability of content restoration based on an adaptive bit allocation mechanism was proposed in [22]. In this scheme, the restoration-bits for tampering recovery were produced according to the priority of each block by using the non-subsampled contourlet transform (NSCT) coefficients. Due to the low embedding volume, the visual quality of the watermarked image is high. Qian et al. [25] proposed an image self-embedding algorithm, in which the cover image was compressed into a number of bits by multi-level encoding. As a result, each block was encoded into 64 bits on average. On the receiver side, after de-quantization, inverse DCT, and the rounding operation, the reference-bits were decompressed, and the tampered blocks can be recovered. Instead of embedding gray-level information or frequency coefficients, Yang et al. [26] created an index table of the cover image via vector quantization (VQ) and generated a pseudo-random sequence to determine where to embed the VQ indices of all of the blocks. If the watermarked image was tampered, the VQ index table can be reconstructed and used to restore the tampered regions by the VQ codewords. However, if all of the embedded copies of the VQ index of the block were destroyed, the quality of the recovered image using this method was not high enough.

In this work, we propose a fragile watermarking scheme for recovering tampered images based on vector quantization and image inpainting. In order to enhance the recovery quality, the merits of the two techniques, i.e., VQ and inpainting, are integrated in the generation of recovery-bits for each classified cover image block. For the blocks with the complex distribution characteristics, the VQ index is used to produce the recovery-bits, whereas, for the smooth blocks, both the VQ index and the inpainting indicator are used to form the recovery-bits. On the receiver side, an adaptive mechanism is utilized to recover all localized, tampered blocks after judging the block classification through the extracted recovery-bits. Because the validity of the extracted VQ indices and the damaged degree of the neighboring regions are considered, the recovery procedure can choose the appropriate method adaptively, i.e., VQ or image inpainting, to recover each tampered block and achieve high quality of the recovered image.

The rest of the paper is organized as follows. Section 2 describes the watermark embedding procedure of the proposed scheme, Section 3 presents the procedure of adaptive tampering recovery, experimental results and comparisons are given in Section 4, and Section 5 concludes the paper.