Techniques and methodologies for validating the authenticity of digital images and testing for the presence of doctoring and manipulation operations on them has recently attracted attention. We review three categories of forensic features and discuss the design of classifiers between doctored and original images. The performance of classifiers with respect to selected controlled manipulations as well as to uncontrolled manipulations is analyzed. The tools for image manipulation detection are treated under feature fusion and decision fusion scenarios.

Biometric applications are stepping into a new dimension with respect to size and complexity, due to the decision of many countries to introduce biometric travel documents. Thus, many problems become increasingly relevant, two of which we address: secure and privacy-preserving storage of biometric references and issues of sensor interoperability. For the first problem, a review of alternative storage methods (physical, optical, centralized database) is presented under consideration of security aspects. The second problem is addressed by a review of scientific activities in the domain and a detailed discussion for one specific exemplary biometric: online handwriting. Results of an experimental intersensor cross-validation are presented, which consist of different semantic classes, forgery strengths, and sensor characteristics. The main conclusions are that the degradation of recognition acuracy in online signature verification may reach up to a factor of 4.4 for nonskilled forgeries in cross-sensor verification scenarios, but is not as significant for skilled forgeries and is of the same order as for fingerprint systems. Further, the enrollment data representation of the chosen reference algorithm, BioHash, is compact enough for storage in identity (ID) documents. Thus, the application of online handwriting as supplementary modality appears feasible in future interoperable ID document scenarios.

We investigate the performance of state of the art universal steganalyzers proposed in the literature. These universal steganalyzers are tested against a number of well-known steganographic embedding techniques that operate in both the spatial and transform domains. Our experiments are performed using a large data set of JPEG images obtained by randomly crawling a set of publicly available websites. The image data set is categorized with respect to size, quality, and texture to determine their potential impact on steganalysis performance. To establish a comparative evaluation of techniques, undetectability results are obtained at various embedding rates. In addition to variation in cover image properties, our comparison also takes into consideration different message length definitions and computational complexity issues. Our results indicate that the performance of steganalysis techniques is affected by the JPEG quality factor, and JPEG recompression artifacts serve as a source of confusion for almost all steganalysis techniques.

We describe a novel framework for watermarking 3-D objects using their texture or silhouette information. Unlike most conventional 3-D object-watermarking techniques, for which both insertion and extraction of the mark are performed on the object itself (3-D/3-D approach), we propose an asymmetric 3-D/2-D procedure. It consists in watermarking 3-D objects and retrieving the mark from rendering 2-D images or videos having used the 3-D synthetic object, thus protecting the visual representations of the object. Two 3-D object watermarking schemes are presented, a texture-based approach for 3-D photorealistic objects and a silhouette-based approach for 3-D CAD (computer-assisted design) objects.

While digital watermarking has received much recent attention, it is still a relatively young technology. As such, there are few accepted tools and metrics that can be used to validate the performance claims asserted by members of the research community and evaluate the suitability of a watermarking technique for specific applications. This lack of a universally adopted set of metrics and methods has motivated us to develop a web-based digital watermark evaluation system known as the watermark evaluation testbed or WET. This system is designed to be focused on still image digital watermarking. WET consists of reference software that includes both watermark embedders and detectors, attacks, evaluation modules, and a large image database. This system has undergone several improvements since its inception. Our ultimate goal has been to develop a platform where any watermarking researcher can test not only the performance of known techniques, but also their own techniques. This goal is reached by the latest version of the system. New tools and concepts are designed to achieve the desired objectives. Moreover, we also summarize the development process of the entire project as well as introduce new directions for future work. ©

Robust still image watermarks are evaluated in terms of image fidelity and robustness. We extend this framework and apply reliability testing to robust still image watermark evaluation. Reliability is the probability that a watermarking algorithm will correctly detect or decode a watermark for a specified fidelity requirement under a given set of attacks and images. In reliability testing, a system is evaluated in terms of quality, load, capacity, and performance. To measure quality that corresponds to image fidelity, we compensate for attacks to measure the fidelity of attacked watermarked images. We use a conditional mean of pixel values to compensate for valumetric attacks such as gamma correction and histogram equalization. To compensate for geometrical attacks, we use error concealment and perfect motion estimation assumption. We define capacity to be the minimum embedding strength parameter and the maximum data payload that meet a specified error criteria. Load is then defined to be the actual embedding strength and data payload of a watermark. To measure performance, we use bit error rate (BER) and receiver operating characteristics (ROCs) of a watermarking algorithm for different attacks and images. We evaluate robust watermarks for various loads, attacks, and images.

We explore how an RGB representation can be computed from a spectral description of real images, with many different colors. To pass from spectral distribution to tristimulus values, several color-matching functions (CMFs) were proposed, derived from experimental setups with simplified visual conditions, considering the colors pointwise and independently from any visual context. A high interdifference is observed in cone spatial distribution between human subjects, without any corresponding significant difference in final color sensation. It is likely that a spatial compensation is performed by human observers that strongly decreases the subjectivity in color perception, and we ask if a similar principle should be considered in digital imaging. We investigate the interdifference among some CMFs when used to compute color information in complex visual conditions, where multiple, spatially distributed different colors are present. This is relevant in synthetic image generation of scenes under different illuminants, computed at the spectral energy distribution at 5-nm intervals, to be converted into RGB for monitor display. The analysis of the interdifference among tristimulus colors obtained using different CMFs shows a significant decrease of the interdifference when a contextual color correction is applied, based on Von Kries or Retinex methods

We develop a methodology to find the optimal memory color (colors of familiar objects) boundary in YCbCr color space and a local image adjustment technique called preferred color reproduction (PCR) to improve image quality. The optimal memory color boundary covers most familiar object colors taken under various viewing conditions. The PCR algorithm is developed based on the idea that colors of familiar objects (memory colors) are key factors in judging the naturalness of an image. The PCR algorithm is applied to pixels detected as having a memory color. Memory color detection is conducted using color information by checking if an input color is inside the predetermined memory color boundary. The PCR algorithm transforms colors inside the memory color boundary to be shifted toward the boundary of constant interval in the center. The PCR algorithm is applied to skin colors, and psychophysical experiments using real images were conducted to determine the best parameters for the algorithm resulting in the most preferred image.

Two methods for colorimetric characterization of color scanner are proposed based on the measures of perceptual color difference error. The first method is used to minimize the total color differences between the actual and predicted color samples. The second one, which is a generalization of the existing cubic-root preprocessing technique, derives the mapping between the p'th root of scanner responses and Commission Internationale de l'Eclairage L*a*b* (CIELAB) values. The experiment results indicate that the color accuracies of the proposed methods, especially the second one, are better than those of the traditional CIE XYZ (CIEXYZ)-space-based characterization methods.

We develop a wavelet-based codec using a low-cost and low-energy-consuming 16-bit fixed-point digital signal processor (DSP). The target application is designed to grab an image from a camera, code the image using a wavelet-based codec, and send a coded file over a wireless global system for mobile communications/general packet radio services (GSM/GPRS) network. Since the channel capacity of the GSM/GPRS network is limited, a DSP with a relatively low computational power is sufficient to implement an image codec. A trade-off is made between the complexity and the efficiency of the wavelet-based image coder. Any implementation issues concerning the wavelet transform and entropy coding are discussed. Lifting schemes for a 9-7 and a 5-3 filter bank are used. The novel ordering of bit-plane bits is presented. The proposed codec is comparable with a JPEG2000 codec in the rate-distortion sense. The experimental results show that the implementation of JPEG 2000 on the presented platform runs 60% slower than the proposed codec. The coded file is transmitted over a wireless network using Internet protocol/user datagram protocol/trivial file transfer protocol (IP/UDP/TFTP).

The quantitative description of the size of a display gamut is crucial for designing the wide-gamut display of high saturation primaries. A method to count the discernible colors in a display gamut with the color-difference formula CIE94 is shown. The color difference between neighboring discernible colors in a display gamut is assumed to be unit E94 (E94=1). The counting result shows that there are 199,491 discernible colors for the ITU-R BT.709 standard display. The number of discernible colors of two wide-gamut LED displays is compared with the standard display, which is more meaningful for representing their relative gamut sizes. For the considered three- and four-primary LED displays, their gamut volumes in CIELAB color space are respectively 1.76 and 1.98 times larger than the standard display. It is found that the number of discernible colors of the two LED displays is only 1.38 and 1.49 times larger than the standard display. The impact of flare on the reduction of the number of discernible colors and gamut volume is shown. The number of discernible colors and gamut volume is respectively reduced by 30%, with about 3.3 and 2.5% flare levels for the three displays considered in this work.

An effective algorithm to reduce gray-level disturbance (GLD) in pulse number modulation is proposed. GLD occurs when moving image sequences are presented by plasma display panels (PDPs) or digital micromirror devices (DMDs), which use pulse number modulation to express gray levels. We first develop a systematic model for GLD, and then show that GLD can be eliminated if the light emission pattern of every gray level has the same shape. Based on the ideal condition, we design subfield and driving vectors. The lexicographically largest vector is employed as the subfield vector, since it can flexibly control the shapes of light emission patterns. Then, three methods are proposed to determine driving vectors: the zero-order, first-order, and tree methods. The zero-order method has the lowest implementation complexity, whereas the tree method reduces GLD most effectively. The first-order method offers a good tradeoff between implementation complexity and disturbance reduction capability. Simulation results demonstrate that these methods suppress GLD effectively and provide good moving image quality.

A deinterlacing algorithm based on edge-dependent interpolation (EDI) that considers edge patterns is proposed. Generally, EDI algorithms perform visually better than other deinterlacing algorithms using one field. However, they produce unpleasant results due to failure in estimating edge direction. To estimate the edge direction precisely, not only simple differences between adjacent two lines but also edge patterns are used. Edge patterns, which give sufficient information to estimate the edge direction, appear along the edge direction. Therefore, we analyze properties of edge patterns and model them as a weight function. The weight function helps the proposed method to estimate the edge direction precisely. Experimental results indicate that the proposed algorithm outperforms conventional EDI approaches with respect to both objective and subjective criteria.

The deficiency of the existing level-set-based denoising techniques is that they are sensitive to noise. This is due to that fact that the curvature and gradient measurements in the partial differential equation are very sensitive to noise, and the denoising performance is affected. This work proposes to perform the level-set-based curve evolution on the dyadic wavelet transform domain. The main advantage is that in the dyadic wavelet transform domain, noise has less influence on curvature and gradient measurements as the scale increases. Thus, the edge indicator function value can be directly calculated from the dyadic wavelet coefficients rather than from an external force field by convolving the noisy image with a Gaussian filter. For further reducing the noise at the finest scale where noise is dominant, minimum mean-squared-error (MMSE)-based filtering is performed as the first pass of denoising, followed by performing the level-set curve evolution as the second pass of further denoising and enhancement. Experimental results demonstrate that the proposed algorithm generates state of the art denoising results.

This paper is a revision of a paper presented at the SPIE conference on Medical Imaging 2005: Physiology, Function, and Structure from Medical Images, Feb. 2005, San Diego, California. The paper presented there appears (unrefereed) in SPIE Proceedings Vol. 5746. Segmentation, or separating an image into distinct objects, is the key to creating 3-D renderings from serial slice images. This is typically a manual process requiring trained persons to tediously outline and isolate the objects in each image. We describe a template-based semiautomatic segmentation method to aid in the segmentation process and 3-D reconstruction of microscopic objects recorded with a confocal laser scanning microscope (CLSM). The simple and robust algorithm is based on the creation of a user-defined object template, followed by automatic segmentation of the object in each of the remaining image slices. The user guides the process by selecting the initial image slice for the object template, and labeling the object of interest. The algorithm is applied to mathematically defined shapes to verify the performance of the software. The algorithm is then applied to biological samples, including neurons in the common housefly. It is the quest to further understand the visual system of the housefly that provides the opportunity to develop this segmentation algorithm. Further application of this algorithm may extend to other registered and aligned serial section datasets with high contrast objects.

Setting of initial contours is a critical problem of active contour models. Although gradient vector flow (GVF) has simplified them greatly compared to the traditional snake, additional work is still necessary. We propose a specific setting method based on vector field analysis. Some critical points in the force field are used to estimate curve deformation. These points seem like sources whose neighbor vectors emanate from them, therefore they can be regarded as inflated centers of deformable contours in GVF snakes. The proposed method detects these points and sets initial contours to contain all the ones in the object area and none of the others. It can prevent the contours from stabilizing on false boundaries in homogeneous regions or real boundaries of other objects. Several experimental results demonstrate that this method is robust and effective.

We present a new approach for computer-generated integral photography (IP) based on ray tracing, for the reconstruction of high quality photorealistic 3-D images of increased complexity. With the proposed methodology, all the optical elements of a single-stage IP capturing setup are physically modeled for the production of real and virtual orthoscopic IP images with depth control. This approach is straightforward for translating a computer-generated 3-D scene to an IP image, and constitutes a robust methodology for developing modules that can be easily integrated in existing ray tracers. An extension of this technique enables the generation of photorealistic 3-D videos [integral videography (IV)] and provides an invaluable tool for the development of 3-D video processing algorithms.

This paper is a revision of a paper presented at the SPIE conference on Visual Communications and Image Processing 2003, Jul. 2003, Lugano, Switzerland. The paper presented there appears (unrefereed) in SPIE proceedings Vol. 5150. In near-lossless image coding, each reconstructed pixel of the decoded image differs from the corresponding one in the original image by not more than a prespecified value. Such schemes are mainly based on predictive coding techniques, which are not capable of quality or resolution-wise scalable decoding. Lossless image coding with scalable decoding is mainly based on transforms that map integers to integers using lifting factorization. In this work, the near-lossless quantization is incorporated into lifting to develop a wavelet-based near-lossless image coding scheme that supports scalability. The proposed technique, which performs online quantization, eliminates the inefficiencies of prequantization-based near-lossless coding and the difficulty in wavelet domain near-lossless quantizing. Two online near-lossless quantization techniques based on 1-D and 2-D transforms are presented. The algorithms outperform the prequantization-based near-lossless image coding in both bit rate and root mean square (rms) error performances, resulting in both subjectively and objectively superior performance in scalable decoding. The 2-D online scheme results in comparable performance with JPEG-LS, which is a nonscalable coding technique. Using these novel schemes enables scalable decoding of near-lossless coded images at the expense of a small increase in bit rates compared to those achieved using JPEG-LS.

Color separation and highly optimized context tree modeling for binary layers have provided the best compression results for color map images that consist of highly complex spatial structures but only a relatively few number of colors. We explore whether this kind of approach works on photographic and palette images as well. The main difficulty is that these images can have a much higher number of colors, and it is therefore much more difficult to exploit spatial dependencies via binary layers. The original contributions of this work include: 1. the application of context-tree-based compression (previously designed for map images) to natural and color palette images; 2. the consideration of four different methods for bit-plane separation; and 3. Extension of the two-layer context to a multilayer context for better utilization of the crosslayer correlations. The proposed combination is extensively compared to state of the art lossless image compression methods.

For copyright protection, the robustness of a watermarking scheme against various attacks is an essential requirement. Many proposed robust watermarking schemes may achieve good robustness but at the expense of sacrificing the good quality of the watermarked image. This work, therefore, proposes a transparent robust watermarking scheme, which embeds the watermark (or the secret information) adaptively in the discrete cosine transform (DCT) domain. The proposed scheme is blind as the original image is not needed, and only an image-dependent dual key containing the watermark locations and the scaling factors is required for watermark extraction. This content adaptive replacement embedding technique can guarantee preservation of good visual quality of a watermarked image. The watermarking scheme is shown to be highly robust for a wide range of attacks, including JPEG compression, quantization, additive noise, rotation, cropping, scaling, and filtering. Compared with other existing DCT-domain watermarking methods, the proposed watermarking method can achieve higher robustness performance, while retaining better quality of the watermarked image.

An adaptive blind wavelet-based watermarking scheme using tree mutual differences (ABW-TMD) is proposed by exploiting mutual differences between grouped coefficients of so-called wavelet trees. The ABW-TMD encoder adaptively searches for the bit host difference in such a manner to minimize the embedding error. The proposed ABW-TMD scheme proves its superiority in resisting against various image processing attacks as well as in reducing the computational cost when compared to other schemes. For the same watermark bit length, a 35% increase in peak signal-to-noise ratio (PSNR) can be achieved over what is possible using an earlier method. Also, when fixing the PSNR of the watermarked image, the watermark bit length can be doubled with the ability to extract the watermark, even in the presence of high JPEG/JPEG2000 compression ratios.

We present cryptanalysis of an image encryption scheme, which is based on the base-switching (BS) lossless compression algorithm. The following conclusions are reached: 1. the size of the key space, i.e., the security against brute-force attacks, was greatly overestimated by the designers; and 2. the scheme is not secure against known/chosen-plaintext/ciphertext attacks. A real example is given to show the feasibility of a proposed chosen-plaintext attack. In addition, some other minor problems of the joint compression-encryption scheme are also pointed out.

An efficient secure encryption scheme for JPEG 2000 code streams is proposed. The scheme does not introduce superfluous JPEG 2000 markers in the protected code stream and protects all the information of an image contained in a JPEG 2000 packet body. It can work with any standard ciphers to encrypt the stream data in sequence and is computationally efficient and memory saving.