The quality assessment is a vital routine in videorelated industries such as broadcast service providers. Due to the duration and the excessive number of the video files, case by case assessment of the files by operators is no longer feasible. Therefore, a computer-based video quality assessment mechanism is the only solution. While it is common to measure the quality of a video file at the compression stage by comparing it against the raw data, at later stages no reference video is available for comparison. Therefore, a no-reference (Blind) video quality assessment (NR-VQA) technique is essential. The common NRVQA methods learn a quality metric based on a number of features extracted from video frames or series of adjacent frames. In the training stage, the features are usually required all at once and the outcome is mainly insensitive to the frame order. For instance, most methods return the same quality score if the video is played in the reverse time order. In this work, we propose an in-the-wild NR-VQA method based on recurrent neural networks (RNN), which takes the frame order into account. Indeed, the RNN is responsible to combine frame-level features by preserving their order so as to form a single video quality metric. As the RNN receives the frame-level statistical features in a sequential manner, the method is also oblivious to the frame size and video length (duration). The experiments show comparable or better performance with previous methods on KonVid-1k dataset.
Paper
Full text
No-Reference Video Quality Assessment using Recurrent Neural Networks
Semantic Scholar · Computer Science · 2019
Abstract
The quality assessment is a vital routine in videorelated industries such as broadcast service providers. Due to the duration and the excessive number of the video files, case by case assessment of the files by operators is no longer feasible. Therefore, a computer-based video quality assessment mechanism is the only solution. While it is common to measure the quality of a video file at the compression stage by comparing it against the raw data, at later stages no reference video is available for comparison. Therefore, a no-reference (Blind) video quality assessment (NR-VQA) technique is essential. The common NRVQA methods learn a quality metric based on a number of features extracted from video frames or series of adjacent frames. In the training stage, the features are usually required all at once and the outcome is mainly insensitive to the frame order. For instance, most methods return the same quality score if the video is played in the reverse time order. In this work, we propose an in-the-wild NR-VQA method based on recurrent neural networks (RNN), which takes the frame order into account. Indeed, the RNN is responsible to combine frame-level features by preserving their order so as to form a single video quality metric. As the RNN receives the frame-level statistical features in a sequential manner, the method is also oblivious to the frame size and video length (duration). The experiments show comparable or better performance with previous methods on KonVid-1k dataset.