Perceptual Watermarks for Digital Image and VideoWhat is ‘perceptual’ watermarkPrior knowledgeSlide 4Slide 5Slide 6Perceptual watermark techniquesImage-Independent watermarkBlock diagram of Cox’s schemeImplementationChallenging problemImage-dependent or Image-adaptive watermarkImage-Adaptive watermarkBlock-based DCT approachSlide 15Block diagram of IA-DCT approachWavelet DWT approachSlide 18DetectionTesting IA-DCT without original imageSlide 21ComparisonVideo watermarksKey points in paperPapersThank you for your attentionPerceptual Watermarks for Perceptual Watermarks for Digital Image and VideoDigital Image and VideoECE 738 paper ECE 738 paper presentationpresentationPei QiPei QiECE at UW-MadisonECE at [email protected]@cae.wisc.eduWhat is ‘perceptual’ What is ‘perceptual’ watermarkwatermarkPrior knowledgePrior knowledgePerceptual watermarkPerceptual watermarkPrior knowledgePrior knowledgeAdditive watermarkAdditive watermarkIdeal watermarkIdeal watermark Three principlesThree principles - Transparency or imperceptibility- Transparency or imperceptibility - Robustness- Robustness - Capacity- Capacity Challenging problemChallenging problem - Conflicts- Conflicts - Tradeoff between transparency and - Tradeoff between transparency and robustnessrobustnessPrior knowledgePrior knowledgeHuman visual systemHuman visual system Three propertiesThree properties of the human visual system of the human visual system 1.1. Frequency sensitivityFrequency sensitivity What’s freq. sensitivityWhat’s freq. sensitivity Freq. sensitivity describes the human eye’s sensitivity to sine wave Freq. sensitivity describes the human eye’s sensitivity to sine wave gratings at various freq. Given that the minimum viewing distance is fixed, it’s gratings at various freq. Given that the minimum viewing distance is fixed, it’s possible to determine a static just noticeable difference threshold for each freq. possible to determine a static just noticeable difference threshold for each freq. band.band. JND thresholdJND threshold The JND threshold is such that changes in the frequency content in the The JND threshold is such that changes in the frequency content in the image in the particular frequency band below the threshold are not noticeableimage in the particular frequency band below the threshold are not noticeablePrior knowledgePrior knowledgeHuman visual systemHuman visual system Three propertiesThree properties of the human visual system of the human visual system 2.2. Luminance sensitivityLuminance sensitivity What’s luminance sensitivityWhat’s luminance sensitivity Luminance sensitivity measures the effects of the detectability threshold Luminance sensitivity measures the effects of the detectability threshold of noise on a constant background, which is a nonlinear function and depends of noise on a constant background, which is a nonlinear function and depends on local image characteristics.on local image characteristics. 3. Contrast masking3. Contrast masking Contrast masking allows more dynamic control of the JND threshold Contrast masking allows more dynamic control of the JND threshold levels. Contrast masking refers to the detectability of one signal in the levels. Contrast masking refers to the detectability of one signal in the presence of another signal.presence of another signal.Prior knowledgePrior knowledgeSummarySummary What is our goal to introduce human visual system in watermarking application?What is our goal to introduce human visual system in watermarking application? 1. 1. Determine if a watermark inserted into a image is invisible or not Determine if a watermark inserted into a image is invisible or not 2. We are always trying to insert the maximum strength and maximum length watermarks into an 2. We are always trying to insert the maximum strength and maximum length watermarks into an image, SINCE more watermarks are insertedimage, SINCE more watermarks are inserted - more robust to attacks- more robust to attacks - more likely to be detected- more likely to be detected Make use of properties of human visual system to adjust the watermark so that it’s perfect for both Make use of properties of human visual system to adjust the watermark so that it’s perfect for both robustness and transparencyrobustness and transparency JNDsJNDs JNDs generated from different properties provide the quantized thresholds for embedding JNDs generated from different properties provide the quantized thresholds for embedding watermarks. watermarks. - upper bounds on watermark strength levels- upper bounds on watermark strength levels - upper bounds on watermark length (capacity)- upper bounds on watermark length (capacity) Note: JND thresholds are NOT a fixed value, which depend on different images and approachesNote: JND thresholds are NOT a fixed value, which depend on different images and approachesPerceptual watermark Perceptual watermark techniquestechniquesImage-Independent watermarkImage-Independent watermarkImage-dependent or Image-adaptive Image-dependent or Image-adaptive watermarkwatermarkImage-Independent Image-Independent watermarkwatermarkA typical method (Cox approach)A typical method (Cox approach)Key pointsKey points•Place watermark in perceptually significant components (low Place watermark in perceptually significant components (low frequency) (for robustness)frequency) (for robustness)–Modify by a small amount below Just-noticeable-difference (JND)Modify by a small amount below Just-noticeable-difference (JND)•Use long random vector as watermark to avoid artifactsUse long random vector as watermark to avoid artifacts•Any difference if using other watermark instead (w-b images, logo)Any difference if using other watermark instead (w-b images, logo)(for imperceptibility & robustness)(for imperceptibility & robustness)Embedding Embedding v’v’i i = v= vii + +  v vi i wwii = v = vii (1+ (1+ w wii) ) •Perform DCT on entire image and embed watermark in DCT Perform DCT on entire image and embed watermark in DCT coefficientscoefficients•Choose N=1000 largest AC coeff. and scale {vi} by a random factorChoose N=1000 largest AC coeff. and scale {vi} by a random factorDetectionDetectionBlock diagram of