Content-based image quality assessment of natural scene image distorted by quantization

In this paper, the subjective image quality for different image content is investigated by psychophysical experiments. The experimental images are the parts from natural scenes distorted by integer transform and quantization in H.264 frame. These images are divided into two types based on the scene content, type I and type II. The perceived thresholds and subjective graded scores for different quantization are obtained using forced choice staircase experiments and graded response experiments, respectively. The subjective assessment results showed that the image quality of type I degrades much more than the type II when quantization steps increase, and the preliminary experiment showed that the existed IQA metric, i.e. SSIM, could not predict it well. We also present a content-based image classifier to predict the two image types. The results show good accordance between the classifier and the subjective assessment.