No-Reference Video Quality Assessment by Deep Feature Maps Relations

Blind video quality assessment (BVQA) approaches are meant to analyze the quality of a distorted video perceived by viewers without knowing anything about the pristine video. Numerous deep network-based techniques have been introduced so far. These methods often pool the features obtained for each frame in different ways to generate a video representation and evaluate the quality. A novel approach is presented in this study for obtaining frame-level features to assess quality. In order to accomplish this, we explored the deep feature maps relations as useful information for video quality assessment. The Gram Matrix generated in each layer is analyzed and explored as higher-order quality features using pre-trained networks. The deep feature relations can be considered similar to the covariance matrix, which indicates the correlations between different feature maps. In fact, these correlations reflect the structural information of each frame. After features extraction and pooling, Using an SVR, a quality score is provided. The efficacy of the suggested features is demonstrated by empirical findings and comparable performance to the state-of-the-art techniques is offered.