Blind Quality Estimation by Disentangling Perceptual and Noisy Features in High Dynamic Range Images

Abstract

Abstract-No reference image quality assessment(NR-IQA) on true High Dynamic Range (HDR) images is an unexplored and interesting research topic that has become relevant with the current boom in HDR technology. We first conduct a study on the performance of existing algorithms for NR-IQA, originally developed for Low Dynamic Range images(LDR), on HDR data. Using the results of our investigation, we then propose a new convolutional architecture for NR-IQA achieving state of the art performance in terms of correlation to human judgment on quality. The proposed model disentangles perceptual effects exhibited by the distorted image from the noise present in the image and is capable of extracting perceptually relevant features without the necessity of handcrafting the features. Our architecture predicts the amount and location of the visual errors present in the distorted images without the reference image and performs comparably to the state of the art Full Reference Image Quality assessment(FR-IQA) algorithms.

Hybrid deep emperor penguin classifier algorithm‐based image quality assessment for visualisation application in HDR environments

“…The mean value for the predicted and subjective score are represented as and , respectively. Table 1 tabulates the simulation results of the proposed HDEPO model and these are compared with the existing quality assessment models such as BLIINDS II [17], BRISQUE [18], HIGRADE‐I [20], HIGRADE‐II [20], GM‐LOG [22], OCPP descriptor [23], RF‐based [24], and CNN‐based method [25].…”

Learning Local Distortion Visibility from Image Quality

“…A scheme of the proposed model is shown in Figure 1. The input of our systems are overlapping patches of 32 × 32 pixels, similar to [15,13]. The distortion, E(i, j), is computed directly as in Eq.…”

“…For the architecture of P-net, we make use of a handcrafted layer that we named as augmented input layer [13]. In this layer, in addition to the luminance values of the N × N block, we compute the mean, variance and Mean Subtracted Contrast Normalized (MSCN) image [16].…”

“…Visual quality scores thus implicitly embed latent information on error visibility. In our recent work [13], we have proposed a deep convolutional neural network (CNN) architecture to disentangle the per pixel distortion and what we called the "perceptual resistance", in the context of noreference quality estimation of high dyanamic range compressed pictures. Our results demonstrated that it is possible to effectively estimate these two terms over a broad range of qualities, starting from supra-threshold quality scores.…”

RAVA: Region-Based Average Video Quality Assessment

“…There are many image quality assessment (IQA) methods. For instance, SSIM [ 8 ] and PSNR [ 9 ] are FR IQA methods, [ 10 , 11 ] and NIMA [ 12 ] are NR IQA methods. Since videos are composed of image frames, we consider extending IQA methods to assess video quality.…”