Combining Deep Learning and Feature Engineering for No-Reference Video Quality Assessment

With the ever-growing demand for high-quality video content, methods for evaluating the aesthetic quality of digital videos have become invaluable tools to ensure a satisfying viewing experience for the end-user. Over the years, many accurate and reliable Full-Reference (FR) video quality assessment (VQA) metrics [4, 5, 10, 16, 17] have been developed that assess quality degradation by comparing the processed video to its original source. However, in scenarios where the reference video is unavailable or impractical to obtain, No-Reference (NR) metrics have to be used, which assess quality solely on the distorted video. Despite intensive research, no NR method so far has been found accurate enough to be widely adopted by industry [9].