Enhancing Video Quality Assessment through Semantic Guidance

With the rapid advancement of network and information technologies, user-generated content (UGC) platforms like TikTok have become essential mainstream media channels. However, video quality often varies significantly due to factors during shooting, compression, and transmission, inconveniencing users. While existing methods assess video quality using objective standards, human subjective perception remains crucial. This paper introduces a simple and effective no-reference video quality assessment model that integrates human perception into the complex video generation process to enhance UGC quality evaluation. We develop a trainable FastVQA branch to extract technical features and incorporate a ConvNeXt branch to capture semantic features. These features are combined and processed through a multi-layer perceptron (MLP) network to predict video quality scores. Our findings demonstrate that incorporating semantic features allows the network to better leverage human perception of video content quality, thereby improving the accuracy of technical assessments. This method not only optimizes traditional quality assessment models but also offers a new research direction, highlighting the importance of human perception in automated video quality evaluation.