A Keras and Tensorflow implementation of video quality assessment using deep neural networks is proposed. We propose CNN + LSTM architecture to recognize and synthesize both spatial and temporal artifacts of video impairements respectively. The architecture is shown below and is borrowed from the following paper.
Handcrafted vs Deep Learning Classification for Scalable Video QoE Modeling
Mallesham Dasari§, Christina Vlachou†, Shruti Sanadhya†‡, Pranjal Sahu§, Yang Qiu§, Kyu-Han Kim†, Samir R. Das§ §Stony Brook University, †HPE Labs, ‡Facebook
- We used video samples of 30 seconds each to train the model.
- We first extract individual frames of the video from each video sample and create a numpy array out frames.
- We create numpy array of MOS (i.e, video quality mean opinion score collected from users) that is corresponding to each sample.
- We feed these frames of a video sample to CNNs followed by a series of Time Distributed LSTM.
- Finally, a softmax is used to classify the video sample quality.
- Python3
- Tensorflow
- Keras
- Jupyter Notebook (optional)
Mallesham Dasari, Pranjal Sahu, Yang Qiu