Video quality assessment with deep architecture

Abstract

Abnormal video quality judgment is an important part of monitoring system maintenance. However, it is still challenging to accurately judge possible anomalies. A novel deep model here is proposed to improve exception detection by considering remote context information in the automatic encoder framework. Specifically, C3D and recurrent neural network (RNN) with long short term memory (LSTM) unit will be used to explore the specified features of video sequence and extract advanced spatiotemporal features. In addition, an automatic encoder is used to represent each video frame to explore the hidden details. In addition, high-frequency structure details in gradient images are explored by the dual stream scheme. Consequently, it enhances its performances. Compared with other advanced methods, the experimental results show the effectiveness of our model.