The Analysis of Quality Indicators on Face Recognition in Video Frames

In this paper we examine the effects of various environmental and camera constraints on the accuracy of deep-learning-based face recognition system. More specifically the paper takes a look on the decline of accuracy of the famous ResNet architecture trained on “pretty” images of celebrities, when it is used on stills from video sequences from real-life television broadcast. It introduces a number of different quality indicators (connected to the sharpness of image, the expression of subject and image contrast) and describes their effect on the results of the convolutional network. It attempts to give recommendations on the ideal values for these indicators in order to choose the best candidate from the video sequence on which to perform the recognition algorithm.