Image replica detection system utilizing R-trees and linear discriminant analysis

Abstract

This manuscript introduces a novel system for content-based identification of image replicas. The proposed approach utilizes image resemblance for deciding whether a test image has been replicated from a certain original or not. We formulate replica detection as a classification problem and show that we can optimize efficiency on a per query basis by dynamically solving a reduced multiclass problem. For this purpose, we investigate the effective coupling of multidimensional indexing and machine learning approaches and we aim to achieve replica detection through the training of classifiers with distortions expected in a replica. Visual descriptors are indexed using an R-tree based multidimensional structure for fast image retrieval. Cases unsuccessfully handled by the R-tree are resolved by a multiclass classifier operating on the transformed feature space that results from the application of linear discriminant analysis (LDA) and principal component analysis (PCA). Experimental results show that the proposed system can identify replicas with high accuracy and facilitate a wide range of applications such as copyright protection, content-based monitoring, content-aware multimedia management, etc.

Introduction

Recent technological advances in the area of multimedia content distribution have resulted in a major reorganization of this trade. Valuable digital artworks can be reproduced and distributed arbitrarily, sometimes without any control by their owners. The identification of replicated data is considered an important issue for a number of applications such as copyright infringement, digital rights management, multimedia management using content-aware networks, monitoring and filtering broadcasted content (e.g., tracking of child pornography content), etc. Among the various types of multimedia content, images are a particularly valuable asset and will be the focus of this manuscript. The approaches that have been proposed for robust image identification are watermarking and, recently, image replica detection algorithms.

Watermarking is the technique of imperceptibly embedding information within the host image content [1]. Although watermarking has attracted considerable interest from both industry and academia, it bears certain deficiencies that pose limitations on its use. The requirement of embedding information in a digital image before it is made public, automatically excludes images that are already in the public domain and need to be copyright protected. Another inherent watermarking drawback is the fact that it is an active technique i.e., it modifies the content of the images to be protected. Although these modifications are in general invisible, they do exist and might create problems in certain content categories like medical images, where quality requirements are extremely high.

In order to overcome these inherent watermarking deficiencies, the scientific community started to investigate robust image identification from a content-based perspective. Replica detection, also referred as replica recognition, near-replica detection, perceptual or robust hashing [2], content-based copy detection [3], and multimedia fingerprinting aims at identifying all images that have been reproduced from a source original through the application of intentional or unintentional manipulations. It is based on image similarity and relies on the assumption that images shares plenty of information with their replicas and yet contains enough information to be discriminated from any other non-replica image. The type and severity of manipulations that should be successfully handled by a replica detection system depend on the target application.

The major benefit of such an approach stems from the fact that no additional information should be embedded within the image content, thus eliminating the invisibility constraint inherent to watermarking systems. On the other hand, the fact that the response speed and efficiency of a replica detection scheme is largely affected by the size of the original/reference image dataset, can be considered as the disadvantage of such an approach. All the above make replica detection an important alternative to watermarking that found applications on many types of multimedia data, such as video [4], [3] and audio [5]. Although the problem formulation as described above, bears many similarities with content based image retrieval (CBIR), certain differences do exist, that are detailed in Section 2.

Image replica detection research is still in its early stages, thus only few works addressing tasks identical or slightly different to the one addressed in this manuscript can be found in the literature. In order to tackle the replica detection problem, existing works aim at (a) optimizing the distance function quantifying the perceptual similarity between two images [6], [7], (b) extracting highly representative and informative features for discriminating between replicas and non-replicas [8], [9], or (c) using machine learning techniques and considering the problem as a classification task [10], [11], [12].

In the first case, Qamra et al. [6] present an enhanced perceptual distance function (DPF) which adaptively chooses a different set of features according to their discriminative power. The benefit of this approach is that unlike other schemes that select that same features for all the images, DPF dynamically activate features (with minimum difference) in a pair-wise fashion. In the same direction Kim [7] use the ordinal measure of DCT coefficients as the feature to represent images and the ordinal measures of AC coefficients for measuring distance similarity. A scheme for the optimal selection of a similarity threshold, based on the maximum a posteriori (MAP) criterion, is used to enhance the efficiency of the employed distance function.

Concerning methods that focus on robust features extraction, Ke et al. [9] use PCA-SIFT [13], a local descriptor that has been shown to be more discriminative and compact than the original SIFT [14], and features several characteristics that are ideal for solving the image replica detection problem. Roy and Chang [8] on the other hand, focus on finding a feature space where any two images in the database are well separated from each other. More precisely, the original images are slightly modified in order to increase their mutual separation within the feature space, while taking care that the perceptual difference between the original and the modified image is kept to a minimum.

Finally, in the group of methods that view the problem as a classification task, Maret et al. [10] propose a method where binary classifiers based on support vector machines are constructed for each original image and are independently applied to decide whether a query image is a replica or not. A variation of this system is described in [11] where indexing is used to perform a coarse and rapid selection of the most likely originals and reduce the number of classifiers that need to be applied. In a more recent work [12] the authors improve their method by trying to estimate and efficiently describe the partition of the image space that contains the replicas of a particular original image.

Even though the systems introduced in the aforementioned papers are trying to tackle the same replica detection problem, the proposed solutions, except from the ones proposed in [10], [11], [12], rely mainly on the discriminative power of the extracted features and the effectiveness of the employed distance function. Thus, no particular attention is paid to the fact that having many similarities with a classification problem, image replica detection might benefit from the use of appropriately trained classifiers. In our work we try to take advantage of this fact by searching for an optimal space where the projection of visual features will enable the construction of more discriminant classifiers. The proposed system operates upon a database of stored originals. Its novelty stems from the fact that image similarity is dealt as a classification problem that employs a training scheme and a suitable feature space transformation in order to increase the system robustness. It generates training images based on the types of attacks that the system is designed to cope with, and during the classification process it uses class statistic information to achieve maximum separability between classes.

More specifically, each image is represented by a feature vector and a multidimensional indexing structure based on R-trees [15] is used for indexing these vectors. The “hyper-bounding boxes” employed by the R-tree are selected using an attack-oriented training strategy that aims at modeling all potential attacks that the system is designed to encounter. The structure returns a relatively small set of images (ideally one) that are candidates for being the original of the query image. In order to resolve cases where more than one candidates are returned by the R-tree we introduce the dynamic use of discriminant techniques. Each candidate original and its modified copies are assumed to form a class. Linear discriminant analysis (LDA) [16] is applied in order to yield more discriminant image representations taking into account class information. The resulting representations are expected to be more easily separable, since the reduced number of involved classes facilitates the estimation of a class-discriminant projection space. A classification function is subsequently applied on the projection space for selecting the image corresponding to the original version of the query, if such an image indeed exists. It must be noted that the manuscript is a largely extended and improved version of [17] where the proposed approach was initially presented.

The rest of the manuscript is organized as follows. Section 2 provides a solid definition of image replica detection and outlines its particularities with respect to image retrieval systems. The proposed image replica detection system is described in Section 3. Section 4 describes the experiments conducted and summarizes the performance evaluation results. Concluding remarks are drawn in Section 5.