Corrupted Reference Image Quality Assessment of Denoised Images
Abstract: We propose corrupted reference image quality assessment (CRIQA), a novel foundation for reasoning about image quality and image denoising problems jointly. In order to assess the visual quality of a processed image relative to an ideal reference image (not provided), we predict the full reference image quality assessment (FRIQA) scores of denoised images without having the direct access to the ideal reference image, but with the help of the observed corrupted image, instead. Our simulation studies verify that the CRIQA scores of denoised images indeed agree with the corresponding FRIQA scores; and human subject studies confirm that CRIQA scores are more consistent with the perceived image denoising quality than the NRIQA scores. We demonstrated the usefulness of CRIQA with an application in denoising parameter tuning. Existing system: Consider the problem of image denoising the problem of recovering the latent ideal noise-free signal of maximum visual quality from a given noisy signal. The existing IQAs are unequipped to handle the assessment of the processed data relative to the ideal reference that exists in theory but we lack direct access to.
This paper establishes a novel foundation for reasoning about human vision and image recovery jointly that we call corrupted reference IQA (CRIQA) . Specifically, we propose to predict the FRIQA scores of denoised images (that cannot be computed directly) with the help of the observed corrupted image. Recalling that FRIQA itself predicts the subjective quality of an image, the CRIQA score provides an assessment of the visual quality of the denoised image signal relative to an ideal reference image (that is not observable).
Proposed system: The experiment was carried out over 200 images from the dataset. The computed CRIQA scores were plotted against the FRIQA scores in Figure 5. A 45o line formed by the FRIQA-CRIQA plot indicates that the proposed quality metric scores corresponded almost exactly to the desired FRIQA scores. The Pearson product-moment correlation coefficients for overall 200 images are summarized in Table I (see “no error�). The resultant correlation coefficients between FRIQA and CRIQA are very high, indicating a strong agreement between FRIQA and CRIQA scores. We conclude that CRIQA prediction of FRIQA scores hold for a variety of noise levels, image contents, and denoising algorithms. The proposed CRIQAWSSIM yields a fairly accurate prediction of FRIQA-WSSIM, allowing us to choose a threshold value that is near W-SSIM-optimal. By comparison, the training WSSIM (the FRIQA-W-SSIM score averaged over many training images) typically peaked at a smaller threshold value (corresponding to under-smoothing).
Advantages: The advantages of CRIQA were verified on a subjective test also. We applied the bivariate Skellam shrinkage function in to noisy raw sensor images. We acquired using Sony 6000 and Canon 60D cameras, downsampling by 2_2 to keep the green samples only.
We captured a diverse set of scene contents comprised of natural objects, manmade objects, stuffed animals, human faces, etc. The scene illumination ranged between 66 lux and 30,000 lux. See Figure 8 for examples. The ISO number (200-2500) and exposure settings (1/2500-1/4 seconds) were varied to ensure a wide range of operating conditions. Following the procedures of Figure 7, the parameters used for denoising in were chosen to optimize using MSE, NIQE , dipIQ , SSEQ , W-SSIM, VSNR, resulting in a set of images like the ones shown . Obviously, CRIQA versions of MSE, WSSIM, and VSNR were used because we lack ideal reference images.
Disadvantages: However, in the context of image recovery tasks (such as image denoising), we work only with the data provided by the sensors with no side band information available. Hence RRIQA is incompatible with the image denoising problem. This is not a problem for AWGN corruption (as derivatives can be computed efficiently ,nor for CRIQA metrics (e.g. W-SSIM and VIF) defined in wavelet domain (when the denoising method also operates in the wavelet domain). The above optimization scheme is illustrated .The training-based optimization schemes in and overcomes the pragmatic implementation issues of solving for b and b_ in (1) and (2), respectively. However, unlike the FRIQAoptimal denoising function b_(_), training-optimal denoising function e_(_) is not signal adaptive becauseethod also operates in the wavelet domain).
Modules:
Image denoizing :
Consider the problem of image denoising the problem of recovering the latent ideal noise-free signal of maximum visual quality from a given noisy signal. The existing IQAs are unequipped to handle the assessment of the processed data relative to the ideal reference that exists in theory but we lack direct access to. This paper establishes a novel foundation for reasoning about human vision and image recovery jointly that we call corrupted reference IQA (CRIQA) . Specifically, we propose to predict the FRIQA scores of denoised images (that cannot be computed directly) with the help of the observed corrupted image. Recalling that FRIQA itself predicts the subjective quality of an image, the CRIQA score provides an assessment of the visual quality of the denoised image signal relative to an ideal reference image (that is not observable). Image quality assessment : SUBJECTIVE quality assessment of digital images has many tangible benefits to the design of imaging systems because “noise” and “artifacts” are best described by aspects of images that appear most unnatural to the human eye. An objective visual image quality assessment (IQA) metric aimed at unsupervised prediction of perceived quality expedites the advancement of imaging systems by replacing the subjective analysis with an automatic one. The studies thus far resulted in three regimes of IQA full reference assessment (FRIQA) that compares the perceived similarity of a given image to the ideal reference image; reduced reference assessment (RRIQA) that retains only a few statistics of the reference image; and no reference assessment (NRIQA) which codifies the conformity to a model of a “good image.” Consider the problem of image denoising the problem of recovering the latent ideal noise-free signal of maximum visual quality from a given noisy signal. The existing IQAs are unequipped to handle the assessment of the processed data relative to the ideal reference that exists in theory but we lack direct access to.
Subjective Analysis : The advantages of CRIQA were verified on a subjective test also. We applied the bivariate Skellam shrinkage function in to noisy raw sensor images. We acquired using Sony 6000 and Canon 60D cameras, downsampling by 2_2 to keep the green
samples only. We captured a diverse set of scene contents comprised of natural objects, man-made objects, stuffed animals, human faces, etc. The scene illumination ranged between 66 lux and 30,000 lux. The ISO number (200-2500) and exposure settings (1/2500-1/4 seconds) were varied to ensure a wide range of operating conditions. Following the procedures , the parameters used for denoising in were chosen to optimize using MSE, NIQE , dipIQ, SSEQ , W-SSIM, VSNR, resulting in a set of images like the ones. Obviously, CRIQA versions of MSE, WSSIM, and VSNR were used because we lack ideal reference images. The human observers involved in the subjective assessment were drawn from engineers in camera sensor industry, and the graduate students majoring in electrical and computer engineering or computer science. They were first presented with a 608_480 pixel cropping of the noisy input image . full-HD display at the full pixel resolution.