Linking Bitstream Information to QoE: A Study on Still Images Using HEVC Intra Coding

Predicting individual quality ratings of compressed images through deep CNNs-based artificial observers

“…Deep CNNs have been largely used in media quality assessment despite the lack of large scale subjectively annotated datasets necessary for an effective training process [31,32,33,34,35,36]. To overcome the lack of training samples, many authors relied on transfer learning approaches [37,38] and data augmentation methods [39,40,41,42]. The computer vision community has developed a number of data augmentation approaches [43].…”

“…In [39,40], in addition to the subjectively annotated training samples, the authors created new training samples for which they computed objective measures to be used as a substitute of the MOS. The authors in [41,42], instead, proposed an approach to combine different subjectively annotated datasets into a single larger one, thus overcoming the issues stemming from the different contexts in which the subjective experiments have been conducted. In particular, the MOS values have to be realigned to take into consideration the context influence factors that may affect the result of each experiment in several different ways [44].…”

“…2. Unlike the previous data augmentation approaches designed for media quality assessment that combine different subjectively annotated datasets [41,42] or annotate the stimulus with the scores predicted by objective metrics [39,40], this work exploits a small scale subjectively annotated dataset to figure out an annotation rule that is then used to synthetically create a large number of training samples. It is worth noting here that such a number is several decades larger than what can be subjectively annotated.…”