An Image Quality Predictor based on Convolution Neural Networks

Image quality assessment is an essential issue for image encoding, monitoring, and pricing. We developed an image quality predictor by using convolution neural networks (CNN) in this paper to address this issue. This predictor consists of a distortion classifier and five scorers for predicting the quality of images with different distortions. Different from related work, the proposed predictor uses the Laplacian filter in data preprocessing to increase the gradient variances caused by different distortions and replaces the flatten layer with the global average pooling layer in CNN for reducing the amount of data computation. Our experimental results have shown that the proposed predictor can provide higher prediction accuracy and spend less time on image quality assessment than related work.