Graphics Processing Units (GPU) and their development tools have advanced recently, and industry has become more interested in using them. Among several development frameworks for GPU(s), OpenCL provides a programming environment to write portable code that can run in parallel. This report describes two case studies of algorithm implementations in OpenCL. The first algorithm is Median Filtering which is a widely used image processing algorithm. The other algorithm is RSA which is a popular algorithm used in encryption. The CPU and GPU implementations of these algorithms are compared in method and speed. The GPU implementations are also evaluated by efficiency, stability, scalability and portability. We find that the GPU implementations perform better overall with some exceptions. We see that a pure GPU solution is not always the best and that a hybrid solution with both CPU and GPU may be to prefer in some cases.

Median filtering is a non-linear filtering technique which is effective in removing impulsive noise from data. In this thesis, directional median filtering has been implemented using cumulative histogram of samples in several directions. Different methods to implement directional median filtering have been proposed. The filtered images are smoothed along the direction of the filtering window. All implementations aimed to generate outputs in the least amount of time, while reducing the resource utilization on hardware. The implementation methods were designed for Xilinx Virtex 5 FPGA devices but were also attempted on Spartan 3E. The proposed methods used less than 30% of the resources on Virtex 5 FPGA but the resource utilization on Spartan 3E exceeded the number of available resources. After an initial delay, methods 1 and 2 generate a new output for every 5 clock cycles while method 3 generates an output for every 1.5 clock cycles.

La criminalistique en images numériques se développe comme un outil puissant pour l'authentification d'image, en travaillant de manière passive et aveugle sans l'aide d'informations d'authentification pré-intégrées dans l'image (contrairement au tatouage fragile d'image). En parallèle, l'anti-criminalistique se propose d'attaquer les algorithmes de criminalistique afin de maintenir une saine émulation susceptible d'aider à leur amélioration. En images numériques, l'anti-criminalistique partage quelques similitudes avec la restauration d'image : dans les deux cas, l'on souhaite approcher au mieux les informations perdues pendant un processus de dégradation d'image. Cependant, l'anti-criminalistique se doit de remplir au mieux un objectif supplémentaire, extit{i.e.} : être non détectable par la criminalistique actuelle. Dans cette thèse, nous proposons une nouvelle piste de recherche pour la criminalistique en images numériques, en tirant profit des concepts/méthodes avancés de la restauration d'image mais en intégrant des stratégies/termes spécifiquement anti-criminalistiques. Dans ce contexte, cette thèse apporte des contributions sur quatre aspects concernant, en criminalistique JPEG, (i) l'introduction du déblocage basé sur la variation totale pour contrer les méthodes de criminalistique JPEG et (ii) l'amélioration apportée par l'adjonction d'un lissage perceptuel de l'histogramme DCT, (iii) l'utilisation d'un modèle d'image sophistiqué et d'un lissage non paramétrique de l'histogramme DCT visant l'amélioration de la qualité de l'image falsifiée; et, en criminalistique du filtrage médian, (iv) l'introduction d'une méthode fondée sur la déconvolution variationnelle. Les résultats expérimentaux démontrent l'efficacité des méthodes anti-criminalistiques proposées, avec notamment une meilleure indétectabilité face aux détecteurs criminalistiques actuels ainsi qu'une meilleure qualité visuelle de l'image falsifiée par rapport aux méthodes anti-criminalistiques de l'état de l'art. / Image forensics enjoys its increasing popularity as a powerful image authentication tool, working in a blind passive way without the aid of any a priori embedded information compared to fragile image watermarking. On its opponent side, image anti-forensics attacks forensic algorithms for the future development of more trustworthy forensics. When image coding or processing is involved, we notice that image anti-forensics to some extent shares a similar goal with image restoration. Both of them aim to recover the information lost during the image degradation, yet image anti-forensics has one additional indispensable forensic undetectability requirement. In this thesis, we form a new research line for image anti-forensics, by leveraging on advanced concepts/methods from image restoration meanwhile with integrations of anti-forensic strategies/terms. Under this context, this thesis contributes on the following four aspects for JPEG compression and median filtering anti-forensics: (i) JPEG anti-forensics using Total Variation based deblocking, (ii) improved Total Variation based JPEG anti-forensics with assignment problem based perceptual DCT histogram smoothing, (iii) JPEG anti-forensics using JPEG image quality enhancement based on a sophisticated image prior model and non-parametric DCT histogram smoothing based on calibration, and (iv) median filtered image quality enhancement and anti-forensics via variational deconvolution. Experimental results demonstrate the effectiveness of the proposed anti-forensic methods with a better forensic undetectability against existing forensic detectors as well as a higher visual quality of the processed image, by comparisons with the state-of-the-art methods.

This diploma thesis deals with the methods of despeckling in ultrasound images. Ultrasound imaging and related artifacts are described in more details. Ultrasound imaging has its pros and cons, where speckle noise is a disadvantage to be solved. Models of origin of this specific noise are referred too. Practical part of this thesis aims on filtering speckled images by basic and advanced filtering methods as are linear filtering, median filtering, application of Frost filter, QGDCT, geometric filtering, anisotropic diffusion filtering and filtering based on wavelet transformation. Results are compared on the basis of objective criteria.