No-reference quality assessment of H.264/AVC encoded video based on natural scene features

H.264/AVC coded video quality is crucial for evaluating the performance of consumer-level video camcorders and mobile phones. In this paper, a DCT-based video quality prediction model (DVQPM) is proposed to blindly predict the quality of compressed natural videos. The model is frame-based and composed of three steps. First, each decoded frame of the video sequence is decomposed into six feature maps based on the DCT coefficients. Then five efficient frame-level features (kurtosis, smoothness, sharpness, mean Jensen Shannon divergence, and blockiness) are extracted to quantify the distortion of natural scenes due to lossy compression. In the last step, each frame-level feature is averaged across all frames (temporal pooling); a trained multilayer neural network takes the five features as inputs and outputs a single number as the predicted video quality. The DVQPM model was trained and tested on the H.264 videos in the LIVE Video Database. Results show that the objective assessment of the proposed model has a strong correlation with the subjective assessment.