We consider image transformation problems, where an input image is transformed into an output image. Recent methods for such problems typically train feed-forward convolutional neural networks using a per-pixel loss between the output and ground-truth images. Parallel work has shown that high-quality images can be generated by defining and optimizing perceptual loss functions based on high-level features extracted from pretrained networks. We combine the benefits of both approaches, and propose the use of perceptual loss functions for training feed-forward networks for image transformation tasks. We show results on image style transfer, where a feed-forward network is trained to solve the optimization problem proposed by Gatys et al. in real-time. Compared to the optimization-based method, our network gives similar qualitative results but is three orders of magnitude faster. We also experiment with single-image super-resolution, where replacing a per-pixel loss with a perceptual loss gives visually pleasing results.

/pdf/perceptual-losses-for-real-time-style-transfer-and-super-hq4q54uf80.pdf

Perceptual Losses for Real-Time Style Transfer and Super-Resolution

We consider image transformation problems, where an input image is transformed into an output image. Recent methods for such problems typically train feed-forward convolutional neural networks using a \emph{per-pixel} loss between the output and ground-truth images. Parallel work has shown that high-quality images can be generated by defining and optimizing \emph{perceptual} loss functions based on high-level features extracted from pretrained networks. We combine the benefits of both approaches, and propose the use of perceptual loss functions for training feed-forward networks for image transformation tasks. We show results on image style transfer, where a feed-forward network is trained to solve the optimization problem proposed by Gatys et al in real-time. Compared to the optimization-based method, our network gives similar qualitative results but is three orders of magnitude faster. We also experiment with single-image super-resolution, where replacing a per-pixel loss with a perceptual loss gives visually pleasing results.

A survey of skin-color modeling and detection methods

Perceptual visual quality metrics: A survey

Changing network conditions pose severe problems to video streaming in the Internet. HTTP adaptive streaming (HAS) is a technology employed by numerous video services that relieves these issues by adapting the video to the current network conditions. It enables service providers to improve resource utilization and Quality of Experience (QoE) by incorporating information from different layers in order to deliver and adapt a video in its best possible quality. Thereby, it allows taking into account end user device capabilities, available video quality levels, current network conditions, and current server load. For end users, the major benefits of HAS compared to classical HTTP video streaming are reduced interruptions of the video playback and higher bandwidth utilization, which both generally result in a higher QoE. Adaptation is possible by changing the frame rate, resolution, or quantization of the video, which can be done with various adaptation strategies and related client- and server-side actions. The technical development of HAS, existing open standardized solutions, but also proprietary solutions are reviewed in this paper as fundamental to derive the QoE influence factors that emerge as a result of adaptation. The main contribution is a comprehensive survey of QoE related works from human computer interaction and networking domains, which are structured according to the QoE impact of video adaptation. To be more precise, subjective studies that cover QoE aspects of adaptation dimensions and strategies are revisited. As a result, QoE influence factors of HAS and corresponding QoE models are identified, but also open issues and conflicting results are discussed. Furthermore, technical influence factors, which are often ignored in the context of HAS, affect perceptual QoE influence factors and are consequently analyzed. This survey gives the reader an overview of the current state of the art and recent developments. At the same time, it targets networking researchers who develop new solutions for HTTP video streaming or assess video streaming from a user centric point of view. Therefore, this paper is a major step toward truly improving HAS.

A Survey on Quality of Experience of HTTP Adaptive Streaming

Experimental data are presented that clearly demonstrate the scope of application of peak signal-to-noise ratio (PSNR) as a video quality metric. It is shown that as long as the video content and the codec type are not changed, PSNR is a valid quality measure. However, when the content is changed, correlation between subjective quality and PSNR is highly reduced. Hence PSNR cannot be a reliable method for assessing the video quality across different video contents.

Scope of validity of PSNR in image/video quality assessment

Peak Signal-to-Noise Ratio (PSNR) is widely used as a video quality metric or performance indicator. Some studies have indicated that it correlates poorly with subjective quality, whilst others have used it on the basis that it provides a good correlation with subjective data. Existing literature seems to provide conflicting evidence of the accuracy of PSNR as a video quality metric. Based on experimental results, we explain a scenario where PSNR provides a reliable indication of the variation of subjective video quality and scenarios where PSNR is not a reliable video quality metric. We show that PSNR follows a monotonic relationship with subjective quality in the case of full frame rate encoding when the video content and codec are fixed. We provide evidence that PSNR becomes an unreliable and inaccurate quality metric when several videos with different content are jointly assessed. Furthermore, PSNR is inaccurate in measuring video quality of a video content encoded at different frame rates because it is not capable of assessing the perceptual trade-off between the spatial and temporal qualities. Finally, where PSNR is not a reliable video quality metric across different video contents and frame rates, we show that a perceptual video model recently approved by the International Telecommunication Union (ITU) provides quality predictions highly correlating with subjective scores even if different video scenes coded at different frame rates are considered in the test set.

The accuracy of PSNR in predicting video quality for different video scenes and frame rates

With the constant evolution of video technology and the deployment of new video services, content providers and broadcasters always face the challenge of delivering an adequate video quality which meets end-users expectations. The development of reliable quality testing and quality monitoring tools that can be used by broadcasters ultimately requires reliable objective video quality metrics. In turn, the validation of these objective models requires reliable subjective assessment, the most accurate representation of the quality perceived by end-users. Many different subjective assessment methodologies exist, and each has its advantages and drawbacks. One important element in a subjective testing methodology is the choice of the rating scale. In this paper, we make a direct comparison between four scales, which are either included in existing international standards or proposed to be used in future standardization activities related to video quality. We examine the subjective data from the points of view of response behavior from participants, similarity and variability of subjective scores. We discuss these results within the context of the subjective quality assessment of high-definition video compressed and transmitted over error-prone networks. Our experimental data show no overall statistical differences between the different scales. Results also show that the single-stimulus presentation provides highly repeatable results even if different scales or groups of participants are used.

Study of Rating Scales for Subjective Quality Assessment of High-Definition Video

In this paper we describe a subjective quality assessment experiment conducted to measure the impact of temporal artifacts on video quality and characterize the influence of content motion on perceived quality. We examined the human response to jerkiness and jitter by considering different levels of strength, duration and distribution of the temporal impairments. Using videos with high picture quality, we found that for intermediate and high frame rate values video quality was similar independently from the duration of the frame rate decimation. On the other hand, for very low frame rates, overall video quality decreased as the duration of the impairment increased. The results also show that a reduction of the temporal resolution over the entire video does not necessarily lead to a significant loss of quality. Finally, the results of this study do not confirm the traditional thinking of lower-motion content receiving a higher quality than high-motion content for a given frame rate decimation factor. Using several motion descriptors, we observed that for a given sub-optimum frame rate, perceived quality does not necessarily increase with decreasing motion magnitude. More particularly, we found that perceived quality of head-and-shoulders content is severely affected by frame rate decimation although it is characterized by very low motion. Our results suggest that motion magnitude is not the only factor affecting perception of temporal artifacts.

Temporal Aspect of Perceived Quality in Mobile Video Broadcasting

Three-dimensional video content has attracted much attention in both the cinema and television industries, because 3D is considered to be the next key feature that can significantly enhance the visual experience of viewers. However, one of the major challenges is the difficulty in providing high quality images that are comfortable to view and that also meet signal transmission requirements over a limited bandwidth for display on television screens. The different processing steps that are necessary in a 3D-TV delivery chain can all introduce artifacts that may create problems in terms of human visual perception. In this paper, we highlight the importance of considering 3D visual attention when addressing 3D human factors issues. We provide a review of the field of 3D visual attention, discuss the challenges in both the understanding and modeling of 3D visual attention, and provide guidance to researchers in this field. Finally, we identify perceptual issues generated during the various steps in a typical 3D-TV broadcasting delivery chain, review them and explain how consideration of 3D visual attention modeling can help improve the overall 3D viewing experience.

/pdf/the-importance-of-visual-attention-in-improving-the-3d-tv-i10891qjzd.pdf

Quan Huynh-Thu

Papers

Scope of validity of PSNR in image/video quality assessment

The accuracy of PSNR in predicting video quality for different video scenes and frame rates

Study of Rating Scales for Subjective Quality Assessment of High-Definition Video

Temporal Aspect of Perceived Quality in Mobile Video Broadcasting

The Importance of Visual Attention in Improving the 3D-TV Viewing Experience: Overview and New Perspectives