Скачать презентацию ECE 738 Advanced image processing Instructor Dr Hu Скачать презентацию ECE 738 Advanced image processing Instructor Dr Hu

66d0f9738a9e27531a38ab2451c75cc5.ppt

  • Количество слайдов: 18

ECE 738 Advanced image processing Instructor Dr. Hu Impact analysis of digital watermarking on ECE 738 Advanced image processing Instructor Dr. Hu Impact analysis of digital watermarking on perceptual quality using HVS models Pei Qi ECE at UW-Madison [email protected] wisc. edu 1

Outline v v Backgrounds Perceptual quality metrics Simulation and analysis Conclusions 2 Outline v v Backgrounds Perceptual quality metrics Simulation and analysis Conclusions 2

Backgrounds v A often neglected issue - More focus on robustness v - ignore Backgrounds v A often neglected issue - More focus on robustness v - ignore how to evaluate the quality of the watermarked image Perceptual quality requirement for watermarking approaches Transparency - the distortion introduced by embedding watermarks must be imperceptible v How to evaluate the perceptual quality of image - Subjective Assessments - Objective Assessments 3

Current perceptual quality metrics v Current two types of visual quality measures - Subjective Current perceptual quality metrics v Current two types of visual quality measures - Subjective assessment Visual quality is judged by human observers A. Two alternative, Forced Choice (2 AFC) Key points: 1) A pair of images (original and watermarked) 2) Decide which one has higher quality 3) Statistical analysis (tiny difference) 50% : 50% 4

Current perceptual quality metrics v Current two types of visual quality measures - Subjective Current perceptual quality metrics v Current two types of visual quality measures - Subjective assessment Visual quality is judged by human observer B. Five – Scale Rating System (ITU-R Rec. 500) Observers are request to rate the quality of the watermarked image to the different quality scale. 5

Current perceptual quality metrics v Feature - Decision made by human observer v Disadvantage Current perceptual quality metrics v Feature - Decision made by human observer v Disadvantage of subjective assessments - Expensive, non-repeatable, - Hard to distinguish very small difference between original image and watermarked image v Objective measures needed 6

Current perceptual quality metrics v Objective assessment - Currently popular metric: SNR, PSNR 1. Current perceptual quality metrics v Objective assessment - Currently popular metric: SNR, PSNR 1. Error Function e(x, y) = I(x, y) – I’(x, y) 2. Mean Square Error (MSE) 3. Signal to noise ratio (SNR) 4. Peak Signal to noise ratio (PSNR) Maximum pixel’s luminance value (0~255) 7

Current perceptual quality metrics v Feature 1) Mean Square Error (MSE) based 2) Independent Current perceptual quality metrics v Feature 1) Mean Square Error (MSE) based 2) Independent of images v Disadvantages - Ignore the fact that judging the perceptual quality of image significantly depends on the human observers - Don’t take into account the effect of HVS 8

Requirements for improved measures Ideally, the improved methods are ought to combine merits of Requirements for improved measures Ideally, the improved methods are ought to combine merits of the two types of assessments v Automated, objective measure v Consider Human visual System (HVS) v 9

Two Improved measures v Weighted PSNR (WPSNR) - Based on the fact that human Two Improved measures v Weighted PSNR (WPSNR) - Based on the fact that human eye is less sensitive to changes in textured areas than in smooth areas - WPSNR use an additional parameter called Noise visibility Function (NVF), a texture masking function, as a weight factor Any VNF < 1, WPSNR will be slightly higher than PSNR. Reflect human eye is less sensitive to changes in textured areas. 10

Two Improved measures What’s NVF? - NVF uses a Gaussian model to estimate how Two Improved measures What’s NVF? - NVF uses a Gaussian model to estimate how much texture exists in any area of an image Behavior: - flat regions NVF -> 1 - edge and textured regions NVF -> 0 - Simplified NVF Function Where is the luminance variance for the block of pixels used to estimate NVF and NORM is a normalization function. 11

Two Improved measures v Based on Watson visual models - HVS Consists of JND Two Improved measures v Based on Watson visual models - HVS Consists of JND thresholds (consider the properties of HVS) - Generate a 8 X 8 perceptual threshold matrix of DCT transform - Compare watermarked image with original block by block using perceptual threshold matrix as reference - Compute perceptual error (average of all blocks error) 12

Two Improved measures Compare 8*8 blocks of the watermarked image with the corresponding blocks Two Improved measures Compare 8*8 blocks of the watermarked image with the corresponding blocks of the original image to determine if the block has been modified to the extent that the modification can be perceptible to human observers. Here we compute the average perceptual error of all blocks error Block(i) Perceptual threshold matrix derived from Batson visual models 13

Simulation and Analysis baboon v v v All test images are 8 -bit grayscale Simulation and Analysis baboon v v v All test images are 8 -bit grayscale images of size 256 x 256 pix Cox watermarking approach Original images are: 1) baboon. bmp heavily textured image 2) lena. bmp common image lena 14

Simulation and Analysis General accept threshold 35~40 d. B original a=0. 1 psnr=34. 9 Simulation and Analysis General accept threshold 35~40 d. B original a=0. 1 psnr=34. 9 wpsnr=38. 2 a=0. 15 psnr=32. 5 wpsnr=36. 1 a=0. 1 psnr=33. 6 wpsnr=35. 3 a=0. 15 psnr=30. 3 wpsnr=31. 6 15

Simulation and Analysis 1) When scaling factor fixed, textured image ‘Baboon’ has higher psnr Simulation and Analysis 1) When scaling factor fixed, textured image ‘Baboon’ has higher psnr and wpsnr 2) The wpsrn value is more close to psnr in lena (because of nvf) 3) This indicate that for textured images, we can increase more watermark energe without sacrificing perceptual quality. 16

Simulation and Analysis We can observe 1) Moretextured image ‘baboon’ has much smaller perceptual Simulation and Analysis We can observe 1) Moretextured image ‘baboon’ has much smaller perceptual error than ‘lena’. 2) In other words, image ‘baboon’ is more far from JND threshold, while image ‘pool’ is more likely approach JND threshold. 3) Therefore, ‘baboon’ has higher perceptual capacity 17

Conclusions v v Two improved methods (WPSNR and JND-based) taking into account the effect Conclusions v v Two improved methods (WPSNR and JND-based) taking into account the effect of HVS are more effectively and precisely in evaluating the perceptual quality of watermarked image. However, due to the complexity of HVS, so far we cannot take full advantage of HVS, therefore yell the limitation of using these visual models. Thank you 18