Perceptual Quality Metrics Research Articles

Objective Video Quality Assessment Method for Object Recognition Tasks

In the field of video quality assessment for object recognition tasks, accurately predicting the impact of different quality factors on recognition algorithms remains a significant challenge. Our study introduces a novel evaluation framework designed to address this gap by focussing on machine vision rather than human perceptual quality metrics. We used advanced machine learning models and custom Video Quality Indicators to enhance the predictive accuracy of object recognition performance under various conditions. Our results indicate a model performance, achieving a mean square error (MSE) of 672.4 and a correlation coefficient of 0.77, which underscores the effectiveness of our approach in real-world scenarios. These findings highlight not only the robustness of our methodology but also its potential applicability in critical areas such as surveillance and telemedicine.

DiffRAW: Leveraging Diffusion Model to Generate DSLR-Comparable Perceptual Quality sRGB from Smartphone RAW Images

Deriving DSLR-quality sRGB images from smartphone RAW images has become a compelling challenge due to discernible detail disparity, color mapping instability, and spatial misalignment in RAW-sRGB data pairs. We present DiffRAW, a novel method that incorporates the diffusion model for the first time in learning RAW-to-sRGB mappings. By leveraging the diffusion model, our approach effectively learns the high-quality detail distribution of DSLR images, thereby enhancing the details of output images. Simultaneously, we use the RAW image as a diffusion condition to maintain image structure information such as contours and textures. To mitigate the interference caused by the color and spatial misalignment in training data pairs, we embed a color-position preserving condition within DiffRAW, ensuring that the output images do not exhibit color biases and pixel shift issues. To accelerate the inference process of DiffRAW, we designed the Domain Transform Diffusion Method, an efficient diffusion process with its corresponding reverse process. The Domain Transform Diffusion Method can reduce the required inference steps for diffusion model-based image restoration/enhancement algorithms while enhancing the quality of the generated images. Through evaluations on the ZRR dataset, DiffRAW consistently demonstrates state-of-the-art performance across all perceptual quality metrics (e.g., LPIPS, FID, MUSIQ), while achieving comparable results in PSNR and SSIM.

Single image super-resolution with denoising diffusion GANS

Single image super-resolution (SISR) refers to the reconstruction from the corresponding low-resolution (LR) image input to a high-resolution (HR) image. However, since a single low-resolution image corresponds to multiple high-resolution images, this is an ill-posed problem. In recent years, generative model-based SISR methods have outperformed conventional SISR methods in performance. However, the SISR methods based on GAN, VAE, and Flow have the problems of unstable training, low sampling quality, and expensive computational cost. These models also struggle to achieve the trifecta of diverse, high-quality, and fast sampling. In particular, denoising diffusion probabilistic models have shown impressive variety and high quality of samples, but their expensive sampling cost prevents them from being well applied in the real world. In this paper, we investigate the fundamental reason for the slow sampling speed of the SISR method based on the diffusion model lies in the Gaussian assumption used in the previous diffusion model, which is only applicable for small step sizes. We propose a new Single Image Super-Resolution with Denoising Diffusion GANS (SRDDGAN) to achieve large-step denoising, sample diversity, and training stability. Our approach combines denoising diffusion models with GANs to generate images conditionally, using a multimodal conditional GAN to model each denoising step. SRDDGAN outperforms existing diffusion model-based methods regarding PSNR and perceptual quality metrics, while the added latent variable Z solution explores the diversity of likely HR spatial domain. Notably, the SRDDGAN model infers nearly 11 times faster than diffusion-based SR3, making it a more practical solution for real-world applications.

Differentiable Mean Opinion Score Regularization for Perceptual Speech Enhancement

Many speech enhancement methods require perceptual quality metrics for evaluation. The “holy grail” of perceptual speech quality assessment is human subjective ratings, known as the mean opinion score. However, acquiring human ratings is time-consuming, laborious, and expensive. Existing objective quality metrics, on the other hand, are efficient and easy to compute but do not correlate well with human ratings. In this paper, we propose a relatively lightweight deep-learning-based model to predict the human ratings of speech signals. Since it is differentiable, it can be easily employed as a perceptual regularization to improve existing deep-learning-based speech enhancement methods. Experimental results demonstrate that the predictions of our proposed model correlate well with human judgments. We present application in speech enhancement and show that, interestingly, while there is a degradation in performance in terms of traditional objective metrics, there is a significant improvement in the perceptual quality and the naturalness of the enhanced speech.

Wireless Deep Video Semantic Transmission

In this paper, we design a new class of high-efficiency deep joint source-channel coding methods to achieve end-to-end video transmission over wireless channels. The proposed methods exploit nonlinear transform and conditional coding architecture to adaptively extract semantic features across video frames, and transmit semantic feature domain representations over wireless channels via deep joint source-channel coding. Our framework is collected under the name deep video semantic transmission (DVST). In particular, benefiting from the strong temporal prior provided by the feature domain context, the learned nonlinear transform function becomes temporally adaptive, resulting in a richer and more accurate entropy model guiding the transmission of current frame. Accordingly, a novel rate adaptive transmission mechanism is developed to customize deep joint source-channel coding for video sources. It learns to allocate the limited channel bandwidth within and among video frames to maximize the overall transmission performance. The whole DVST design is formulated as an optimization problem whose goal is to minimize the end-to-end transmission rate-distortion performance under perceptual quality metrics or machine vision task performance metrics. Across standard video source test sequences and various communication scenarios, experiments show that our DVST can generally surpass traditional wireless video coded transmission schemes. The proposed DVST framework can well support future semantic communications due to its video content-aware and machine vision task integration abilities.

High-frequency Normalizing Flow for Image Rescaling.

It is desirable to develop efficient image rescaling methods to transmit digital images with different resolutions between devices and assure visual quality. In image downscaling, the inevitable loss of high-frequency information makes the reverse upscaling highly ill-posed. Recent approaches focus on joint learning of image downscaling and upscaling (e.g., rescaling). However, existing methods still fail to recover satisfactory high-frequency signals when upscaling. To solve it, we propose high-frequency flow (HfFlow), which learns the distribution of high-frequency signals during rescaling. HfFlow is an overall invertible framework with a conditional flow on the high-frequency space to compensate for the information lost during downscaling. To facilitate finding the optimal upscaling solution, we introduce a reference low-resolution (LR) manifold and propose a cross-entropy Gaussian loss (CGloss) to force the downscaled manifold closer to the reference LR manifold and simultaneously fulfill recovering missing details. HfFlow can be generalized to other scale transformation tasks such as image colorization with its excellent rescaling capacity. Qualitative and quantitative experimental evaluations demonstrate that HfFlow restores rich high-frequency details and outperforms state-of-the-art rescaling methods in PSNR, SSIM, and perceptual quality metrics.

CTFlow: Mitigating Effects of Computed Tomography Acquisition and Reconstruction with Normalizing Flows.

Mitigating the effects of image appearance due to variations in computed tomography (CT) acquisition and reconstruction parameters is a challenging inverse problem. We present CTFlow, a normalizing flows-based method for harmonizing CT scans acquired and reconstructed using different doses and kernels to a target scan. Unlike existing state-of-the-art image harmonization approaches that only generate a single output, flow-based methods learn the explicit conditional density and output the entire spectrum of plausible reconstruction, reflecting the underlying uncertainty of the problem. We demonstrate how normalizing flows reduces variability in image quality and the performance of a machine learning algorithm for lung nodule detection. We evaluate the performance of CTFlow by 1) comparing it with other techniques on a denoising task using the AAPM-Mayo Clinical Low-Dose CT Grand Challenge dataset, and 2) demonstrating consistency in nodule detection performance across 186 real-world low-dose CT chest scans acquired at our institution. CTFlow performs better in the denoising task for both peak signal-to-noise ratio and perceptual quality metrics. Moreover, CTFlow produces more consistent predictions across all dose and kernel conditions than generative adversarial network (GAN)-based image harmonization on a lung nodule detection task. The code is available at https://github.com/hsu-lab/ctflow.

Modulation Recognition Using Signal Enhancement and Multistage Attention Mechanism

Robustness against noise is critical for modulation recognition (MR) approaches deployed in real-world communication systems. In MR systems, a corrupted signal is normally enhanced using low-level signal enhancement (SE) before signal classification (SC). Many existing approaches address signal distortion problems by compartmentalizing SE from SC. While those approaches allow for efficient development, they also dictate compartmentalized performance metrics, without feedback from the SC module. For example, SE modules are designed using perceptual signal quality metrics but not with SC in mind. To improve the effectiveness of SE on MR, this paper proposes a joint learning framework consisting of three cascaded modules: dual-channel spectrum fusion, SE, and SC. Instead of separately processing SE and SC, these three modules are integrated into one framework and jointly trained with a single recognition loss. In contrast to estimating clean signals, the SE module in the proposed joint learning framework is trained to predict a ratio mask and find important time-frequency bins for the SC module. We integrate a multistage attention mechanism into the framework to further increase the robustness. The multistage attention mechanism is deployed to strengthen the recognition-related features learned from context information in channel, time, and frequency domains. We evaluate the recognition performance of the proposed framework and its modules on two benchmark datasets: RadioML2016.10a and RadioML2016.10b. The experiment results show that the proposed joint learning framework outperforms the separate learning framework. Moreover, comparisons are performed with several existing learning-based MR methods in the literature. The proposed joint learning framework leads to significant performance improvement, especially for modulated signals corrupted by channel noise.

Compression for Bayer CFA Images: Review and Performance Comparison.

Bayer color filter array (CFA) images are captured by a single-chip image sensor covered with a Bayer CFA pattern which has been widely used in modern digital cameras. In the past two decades, many compression methods have been proposed to compress Bayer CFA images. These compression methods can be roughly divided into the compression-first-based (CF-based) scheme and the demosaicing-first-based (DF-based) scheme. However, in the literature, no review article for the two compression schemes and their compression performance is reported. In this article, the related CF-based and DF-based compression works are reviewed first. Then, the testing Bayer CFA images created from the Kodak, IMAX, screen content images, videos, and classical image datasets are compressed on the Joint Photographic Experts Group-2000 (JPEG-2000) and the newly released Versatile Video Coding (VVC) platform VTM-16.2. In terms of the commonly used objective quality, perceptual quality metrics, the perceptual effect, and the quality-bitrate tradeoff metric, the compression performance comparison of the CF-based compression methods, in particular the reversible color transform-based compression methods and the DF-based compression methods, is reported and discussed.

Downscaling atmospheric chemistry simulations with physically consistent deep learning

Abstract. Recent advances in deep convolutional neural network (CNN)-based super resolution can be used to downscale atmospheric chemistry simulations with substantially higher accuracy than conventional downscaling methods. This work both demonstrates the downscaling capabilities of modern CNN-based single image super resolution and video super-resolution schemes and develops modifications to these schemes to ensure they are appropriate for use with physical science data. The CNN-based video super-resolution schemes in particular incur only 39 % to 54 % of the grid-cell-level error of interpolation schemes and generate outputs with extremely realistic small-scale variability based on multiple perceptual quality metrics while performing a large (8×10) increase in resolution in the spatial dimensions. Methods are introduced to strictly enforce physical conservation laws within CNNs, perform large and asymmetric resolution changes between common model grid resolutions, account for non-uniform grid-cell areas, super-resolve lognormally distributed datasets, and leverage additional inputs such as high-resolution climatologies and model state variables. High-resolution chemistry simulations are critical for modeling regional air quality and for understanding future climate, and CNN-based downscaling has the potential to generate these high-resolution simulations and ensembles at a fraction of the computational cost.

Nonlinear Transform Source-Channel Coding for Semantic Communications

In this paper, we propose a class of high-efficiency deep joint source-channel coding methods that can closely adapt to the source distribution under the nonlinear transform, it can be collected under the name nonlinear transform source-channel coding (NTSCC). In the considered model, the transmitter first learns a nonlinear analysis transform to map the source data into latent space, then transmits the latent representation to the receiver via deep joint source-channel coding. Our model incorporates the nonlinear transform as a strong prior to effectively extract the source semantic features and provide side information for source-channel coding. Unlike existing conventional deep joint source-channel coding methods, the proposed NTSCC essentially learns both the source latent representation and an entropy model as the prior on the latent representation. Accordingly, novel adaptive rate transmission and hyperprior-aided codec refinement mechanisms are developed to upgrade deep joint source-channel coding. The whole system design is formulated as an optimization problem whose goal is to minimize the end-to-end transmission rate-distortion performance under established perceptual quality metrics. Across test image sources with various resolutions, we find that the proposed NTSCC transmission method generally outperforms both the analog transmission using the standard deep joint source-channel coding and the classical separation-based digital transmission. Notably, the proposed NTSCC method can potentially support future semantic communications due to its content-aware ability and perceptual optimization goal.

Do We Need a New Large-Scale Quality Assessment Database for Generative Inpainting Based 3D View Synthesis? (Student Abstract)

The advancement in Image-to-Image translation techniques using generative Deep Learning-based approaches has shown promising results for the challenging task of inpainting-based 3D view synthesis. At the same time, even the current 3D view synthesis methods often create distorted structures or blurry textures inconsistent with surrounding areas. We analyzed the recently proposed algorithms for inpainting-based 3D view synthesis and observed that these algorithms no longer produce stretching and black holes. However, the existing databases such as IETR, IRCCyN, and IVY have 3D-generated views with these artifacts. This observation suggests that the existing 3D view synthesis quality assessment algorithms can not judge the quality of most recent 3D synthesized views. With this view, through this abstract, we analyze the need for a new large-scale database and a new perceptual quality metric oriented for 3D views using a test dataset.

A Compact Noise Covariance Matrix Model for MVDR Beamforming

Acoustic beamforming is routinely used to improve the SNR of the received signal in applications such as hearing aids, robot audition, augmented reality, teleconferencing, source localisation and source tracking. The beamformer can be made adaptive by using an estimate of the time-varying noise covariance matrix in the spectral domain to determine an optimised beam pattern in each frequency bin that is specific to the acoustic environment and that can respond to temporal changes in it. However, robust estimation of the noise covariance matrix remains a challenging task especially in non-stationary acoustic environments. This paper presents a compact model of the signal covariance matrix that is defined by a small number of parameters whose values can be reliably estimated. The model leads to a robust estimate of the noise covariance matrix which can, in turn, be used to construct a beamformer. The performance of beamformers designed using this approach is evaluated for a spherical microphone array under a range of conditions using both simulated and measured room impulse responses. The proposed approach demonstrates consistent gains in intelligibility and perceptual quality metrics compared to the static and adaptive beamformers used as baselines.

Inpainting radar missing data regions with deep learning

Abstract. Missing and low-quality data regions are a frequent problem for weather radars. They stem from a variety of sources: beam blockage, instrument failure, near-ground blind zones, and many others. Filling in missing data regions is often useful for estimating local atmospheric properties and the application of high-level data processing schemes without the need for preprocessing and error-handling steps – feature detection and tracking, for instance. Interpolation schemes are typically used for this task, though they tend to produce unrealistically spatially smoothed results that are not representative of the atmospheric turbulence and variability that are usually resolved by weather radars. Recently, generative adversarial networks (GANs) have achieved impressive results in the area of photo inpainting. Here, they are demonstrated as a tool for infilling radar missing data regions. These neural networks are capable of extending large-scale cloud and precipitation features that border missing data regions into the regions while hallucinating plausible small-scale variability. In other words, they can inpaint missing data with accurate large-scale features and plausible local small-scale features. This method is demonstrated on a scanning C-band and vertically pointing Ka-band radar that were deployed as part of the Cloud Aerosol and Complex Terrain Interactions (CACTI) field campaign. Three missing data scenarios are explored: infilling low-level blind zones and short outage periods for the Ka-band radar and infilling beam blockage areas for the C-band radar. Two deep-learning-based approaches are tested, a convolutional neural network (CNN) and a GAN that optimize pixel-level error or combined pixel-level error and adversarial loss respectively. Both deep-learning approaches significantly outperform traditional inpainting schemes under several pixel-level and perceptual quality metrics.

Dirty Pixels: Towards End-to-end Image Processing and Perception

Real-world, imaging systems acquire measurements that are degraded by noise, optical aberrations, and other imperfections that make image processing for human viewing and higher-level perception tasks challenging. Conventional cameras address this problem by compartmentalizing imaging from high-level task processing. As such, conventional imaging involves processing the RAW sensor measurements in a sequential pipeline of steps, such as demosaicking, denoising, deblurring, tone-mapping, and compression. This pipeline is optimized to obtain a visually pleasing image. High-level processing, however, involves steps such as feature extraction, classification, tracking, and fusion. While this silo-ed design approach allows for efficient development, it also dictates compartmentalized performance metrics without knowledge of the higher-level task of the camera system. For example, today’s demosaicking and denoising algorithms are designed using perceptual image quality metrics but not with domain-specific tasks such as object detection in mind. We propose an end-to-end differentiable architecture that jointly performs demosaicking, denoising, deblurring, tone-mapping, and classification (see Figure 1). The architecture does not require any intermediate losses based on perceived image quality and learns processing pipelines whose outputs differ from those of existing ISPs optimized for perceptual quality, preserving fine detail at the cost of increased noise and artifacts. We show that state-of-the-art ISPs discard information that is essential in corner cases, such as extremely low-light conditions, where conventional imaging and perception stacks fail. We demonstrate on captured and simulated data that our model substantially improves perception in low light and other challenging conditions, which is imperative for real-world applications such as autonomous driving, robotics, and surveillance. Finally, we found that the proposed model also achieves state-of-the-art accuracy when optimized for image reconstruction in low-light conditions, validating the architecture itself as a potentially useful drop-in network for reconstruction and analysis tasks beyond the applications demonstrated in this work. Our proposed models, datasets, and calibration data are available at https://github.com/princeton-computational-imaging/DirtyPixels .

Hybrid Distortion-Based Rate-Distortion Optimization and Rate Control for H.265/HEVC

Mean squared error (MSE) is adopted as the distortion metric for the rate-distortion optimization (RDO) and rate control (RC) in most of the existing video coding standards, e.g., H.265/High-Efficiency Video Coding (HEVC) and H.264/Advanced Video Coding (AVC), because the pixel-wise MSE can represent the image fidelity efficiently. As videos are usually used for visual enjoyment, recently, researchers found that the MSE is not always consistent with perceptual quality, and proposed some perceptual quality-based video coding methods by exchanging the MSE with other perceptual quality metrics completely, like the structure similarity index (SSIM), just noticeable distortion (JND), etc. We believe that a sophisticated video encoder should guarantee both the subject quality and the data fidelity. By investigating the properties of the human visual system (HVS), hybrid distortion-based RDO and RC methods are proposed to improve subjective quality while avoiding fidelity degradation. Experimental results demonstrate that the proposed methods can improve subjective quality efficiently, while preserving the image fidelity of the reconstructed videos as much as possible.

Modeling the Perceptual Quality for Viewport-Adaptive Omnidirectional Video Streaming Considering Dynamic Quality Boundary Artifact

Instead of streaming the entire omnidirectional video that induces a waste of bandwidth, the tile-based viewport adaptive streaming is preferred in practice. Tiles within users’ viewport are transmitted in high quality, while tiles elsewhere are delivered in lower quality to reduce the network bandwidth consumption. In this setup, when users change their viewport, they may potentially be disturbed by the exposure of low-quality tiles and the subsequent quality refinement in a period of delay. We defined this phenomenon in the viewport as a Dynamic Quality Boundary (DQB) artifact, as there exists a clear boundary between low-quality tiles and high-quality tiles which appears, moves, and disappears along with the change of users’ viewport. The DQB artifact is a specific quality degradation in the viewport-adaptive omnidirectional video streaming whose impact on users’ perceptual quality has not been adequately studied. In this paper, we focused on investigating and modeling its impact on the quality perception of users when streaming omnidirectional videos. More specifically, systematic subjective quality evaluation experiments towards DQB artifact were designed and conducted. Accordingly, a perceptual quality metric was proposed considering the quality of tiles, the proportion of low-quality tiles, and the refinement duration. Experimental results, measured in various aspects, showed that the proposed metric can accurately predict users’ perceptual quality of the tile-based viewport-adaptive omnidirectional video streaming service.

Radar Super Resolution Using a Deep Convolutional Neural Network

AbstractSuper resolution involves synthetically increasing the resolution of gridded data beyond their native resolution. Typically, this is done using interpolation schemes, which estimate sub-grid-scale values from neighboring data, and perform the same operation everywhere regardless of the large-scale context, or by requiring a network of radars with overlapping fields of view. Recently, significant progress has been made in single-image super resolution using convolutional neural networks. Conceptually, a neural network may be able to learn relations between large-scale precipitation features and the associated sub-pixel-scale variability and outperform interpolation schemes. Here, we use a deep convolutional neural network to artificially enhance the resolution of NEXRAD PPI scans. The model is trained on 6 months of reflectivity observations from the Langley Hill, Washington, radar (KLGX), and we find that it substantially outperforms common interpolation schemes for 4× and 8× resolution increases based on several objective error and perceptual quality metrics.

Nonlinear Transform Coding

We review a class of methods that can be collected under the name nonlinear transform coding (NTC), which over the past few years have become competitive with the best linear transform codecs for images, and have superseded them in terms of rate-distortion performance under established perceptual quality metrics such as MS-SSIM. We assess the empirical rate-distortion performance of NTC with the help of simple example sources, for which the optimal performance of a vector quantizer is easier to estimate than with natural data sources. To this end, we introduce a novel variant of entropy-constrained vector quantization. We provide an analysis of various forms of stochastic optimization techniques for NTC models; review architectures of transforms based on artificial neural networks, as well as learned entropy models; and provide a direct comparison of a number of methods to parameterize the rate-distortion trade-off of nonlinear transforms, introducing a simplified one.

Filtering Super-Resolution Scan Conversion of Medical Ultrasound Frames

In this paper, we consider a challenging problem of reconstruction of high resolution (HR) B-mode ultrasound (US) image by proposing a novel multi-frame based super-resolution (SR) scan conversion framework. This new framework of SR scan conversion reconstructs an improved HR frame by using the scan data of several low resolution (LR) frames. It also unifies the speckle reduction and HR scan conversion in such a way that it has become a single operation to generate a super-resolved image with lesser loss of information. We evaluated the performance of the proposed model on synthetic images, ultrasound simulated (by Field II software) images and real ultrasound image dataset and the comparison is performed against some of the publicly available state-of-the-art ultrasound image enhancement techniques. Significant improvement in image quality has been achieved due to utilization of non-redundant information present in the scan data of the LR frames. We demonstrate the improvement of the proposed technique through the computation of perceptual and quantitative quality metrics, such as, SSIM, PSNR etc. over the recent competing methods.