Automated Image and Video Quality Assessment for Computational Video Editing

This work is devoted to the development of a complex framework for assessing the quality of images and videos. The objectives of the study are to develop an improved method for image quality assessment and to create a fast and robust method for frame-by-frame scoring of video, which could be also applied to images. This study is dedicated to the non-reference image and video quality assessment methods. The goal of the research is to propose an improvement for an existing machine-learning-based IQA method and adapt it for the task of automated editing of short video clips. The study is focused on the assessment accuracy and performance. Due to this objective three IQA benchmark datasets were downloaded: LIVE, TID2008, and TID2013. Each of them contains a set of reference images in high quality, a set of distorted images generated from the reference set with various distortion types and strength, and subjective score for each image averaged over people surveyed. For the model evaluation on the video processing task, the LIVE Video Quality Challenge dataset was used. DIQA is considered as a baseline method for quality prediction. Proposed enhancements include two patching strategies: uniform patching and object-based patching. Further, an additional step of the hidden layer pre-training with distortion type classification is introduced. Another video scoring metrics are the scale of a scene, face presence in frame and compliance of the shot transitions with the shooting rules The results of this work are applicable to the development of intelligent video and image processing systems, including mobile devices.