Video Quality Assessment Based on the Improved LSTM Model

No‐reference quality assessment of dynamic sports videos based on a spatiotemporal motion model

“…The conditional constraints were incorporated into the objective function of logistic regression. CNN‐MR (CNN and multi‐regression) [14]: The spatial features were captured at the frame level by 2D CNN, and motion information was extracted as temporal information at the sequence level. A multi‐regression model was applied to comprehensively measure video quality. ILSTM (improved LSTM) [16]: To predict video quality, an LSTM was combined with a decision tree method and applied to regression analysis. ConvLSTM (CNN and LSTM): The frame‐level deep features were extracted from the CNN and applied to the LSTM network containing LSTM layers and a fully connected layer. The CNN had 13 convolution layers. ConvGRU (CNN and gated recurrent units): Instead of adopting the LSTM network, a GRU network was used as the regression model with frame‐level deep features.…”

“…The second step is to incorporate the extracted features into the regression framework to make an appropriate NR‐VQA. Several deep learning methods have been investigated for NR‐VQA, for example, CNNs [14], recurrent neural networks (RNNs) [15], long short‐term memory (LSTM) [16], and residual neural networks (ResNet) [17]. Among the various methods, the general‐purpose NR‐VQA method named SACONVA motivates us in this study.…”