Treffer: Analysis and evaluation of No-Reference video quality assesment using neural networks

The main goal of our project is: Analyze the performance of some networks, in order to see if it is capable of classifying, in terms of quality, different types of images. In a very generalized and simplified way, in this project we seek to generate an algorithm that allows us to quantitatively assess the quality of a digital video. This problem will be tackled through the use of self-supervised convolutional neural networks (CNN). This means that our main objective is to be able to generate or train a network whose main source of information is different video chunks, each one with a quality associated with a score. We will obtain this score using a visual perception algorithm called VMAF (Video Multi-Method Assessment Fusion), developed and updated for years by one of the largest video-on-demand and streaming services companies in the world: Netflix. We will mainly use Python and some specialized libraries in other languages such as Bash or C #, for the development of the vast majority of scripts on the data processing steps. From the development of these networks in Keras environment, we will be able to apply different variations to already pre-conceived models existing in the platform. We will use networks and pre-trained models; on which we will make modifications to be able to adapt them to our particular data flow. In our case it will be the VGG16 and ResNet50 networks These networks will be trained in an AWS instance, due to its high computing cost for a conventional GPU. Finally, we will evaluate its performance by analyzing different metrics, such as ¿accuracy¿ and losses in the training process (¿training loss¿). In the final section of conclusions, possible improvements will be discussed and the final analysis of the thesis as a whole will be summarized in a synthesized way.