Subjective video quality assessment (VQA) is the most reliable way to get accurate quality scores, providing first-hand data on the research of quality of experience (QoE). However, traditional subjective VQA has the disadvantage of being cumbersome and time-consuming. In this paper, we propose an efficient subjective VQA framework based on hybrid information and active learning, named HA-SVQA. Built on the principle of active learning for data annotation, HA-SVQA allows iterative assessments on the most valuable or informative videos, which are selected based on hybrid information from the subject’s prior decisions and the objective quality predictions. By eliminating the redundant (or less valuable) videos to be assessed, HA-SVQA can speed up the process of subjective VQA. Concretely, our framework starts with a few quality-known videos to initialize dual-regression models. It then uses a scoring-difference stratified sampling strategy to iteratively select the next group of videos to be assessed, which contain high quality uncertainty. The newly scored videos are used to continually update every part of our framework. By this way, subjective VQA can be stopped early while meeting the acceptable goal of a full-time subjective study. We conducted simulation experiments on three different datasets: LIVE, LIVE VQC, and an underwater video quality dataset. The results show that HA-SVQA can effectively speed up the process of subjective VQA and reduce about 1/3 of the human workload or time cost in the presence of data redundancy. In order to investigate the effectiveness of HA-SVQA in more depth, we conducted a field experiment with deep-sea videos. We found that HA-SVQA is still effective in reducing about 1/3 of the overall human workload, which is consistent with the conclusion of simulation experiments. Finally, we discussed some of the factors that potentially affect the QoE modelling and the subjective VQA.

The no-reference IQA methods based on deep learning have problems of insufficient semantic relevance or high model training requirements. This paper proposes a no-reference IQA based on semantic visual feature tokens and Transformer (VTT-IQA). We firstly use a deep convolutional neural network to extract high-level semantic features of the image, and then map the semantic features to visual feature tokens. Subsequently, the relationship between visual feature tokens is modelled based on the Transformer self-attention mechanism to extract the global information. Meanwhile, a shallow neural network is used to extract the low-level local features of the image and capture its distortion information. Finally, the high- level semantic information and the low-level visual information are integrated to accurately predict the image quality. In order to verify the superiority and robustness of our proposed model, we compared our method with 15 traditional and deep learning based non-reference IQA methods on five mainstream IQA datasets and one underwater IQA dataset, using PLCC and SROCC as the performance evaluation metrics. The experimental results show that the proposed method achieves superior performance with less parameters (about 1.56 MB). Especially, VTT-IQA achieves 0.958 of SROCC on LIVE-MD that contains multiply distorted images. It is proved that VTT-IQA can still accurately evaluate the image quality under complex distortion, and can meet the practical application.

Currently, optical imaging technology has played an important role in deep-sea exploration. However, there is still a lack of research on subjective deep-sea video quality assessment, especially the lack of public deep-sea video quality assessment datasets. We construct a public deep-sea video quality assessment dataset with subjective quality labels, which includes five types of representative real deep-sea scene videos. The original deep-sea video sequences are augmented by two deep-sea video quality enhancement methods that are based on deep learning and fusion respectively, and two video quality degradation methods including Gaussian blurring and Gaussian noise. Subjective video quality assessment is conducted with 20 participants and the absolute category rating method is used for rating. Finally, we obtain a deep-sea video quality assessment dataset with 142 samples. The performance of 8 objective image/video quality assessment models is verified on this dataset. The results show that the current objective video quality assessment models need to be improved for the application in deep-sea video quality assessment. The deep-sea video quality assessment dataset is publicly available at http://ieee-dataport.org/documents/deep-sea-video-quality-dataset. It could help optimize and improve the objective deep-sea video quality assessment models and underwater image/video enhancement technology.

Subjective video quality assessment (SVQA) is essential in order to produce a sufficient benchmark dataset, which is especially important for extreme environments such as underwater videos. However, SVQA has always been plagued by its disadvantages, such as high cost and time-consuming. In this paper, we propose an effective SVQA method based on active learning and clustering (AL-SVQA). Our method initializes dual regression models for video quality prediction with a few quality-known videos. During the quality assessment process with active learning and clustering results, a batch of videos are iteratively selected by a sampling strategy and their quality is scored by a subject. The scored videos are then used to continually update the prediction models until meeting the stop criteria. When the active learning stops, the prediction models will annotate the video scores instead of humans while maintaining good performance. In this way, AL-SVQA can reduce about 40% of human the workload for subjective quality assessment. Evaluation experiments were conducted with an underwater video database.