The amount of collected data for video based applications has increased dramatically in the recent years due to the massive use of video-based systems such as surveillance, broadcasting, product inspection, among others. The huge quantity of videos requires a great number of storage devices and consequently the storage cost may be extremely expensive. Video compression is used to reduce the storage cost by reducing the video storage size without affecting the “user satisfaction”. Thus, compression parameters should be selected based on visual assessment (trying to mimic the visual system). Hence, quality control of video-based systems is a very important task for reducing the storage size by keeping the “user satisfaction”.
Prediction of perceived video quality (PVQ) is an important task for increasing the user satisfaction of video-based systems - not only for evaluating the systems but also for real-time control of, for example, streaming parameters. However, the correlation between PVQ and existing video quality metrics (VQMs) varies across different video content: for one type of scene content, a given metric can be well correlated to humans while at the same time for another type of scene content the correlation is poor. To address the problem of strong VQM dependency on video content, we developed a method to induce content-awareness in the VQMs. Our proposed method is based on analyzing the level of spatial and temporal activity in the video and using those as parameters of an anthropomorphic video distortion model. By doing so, performance of existing VQMs can be increased by up to 20%.