A Data-Driven No-Reference Image Quality Assessment via Deep Convolutional Neural Networks.

Nowadays, image quality assessment (IQA) is important for many applications, such as transmission, compression or reproduction. And no-reference image quality assessment (NRIQA) has received extensive attention because there is no need to use original images. Especially in this artificial intelligence and big data age, it is impossible to get the reference images every time. However, many conventional NR-IQA algorithms extracted features by hand and designed only for one specific distortion type. In this paper, we will propose a NR-IQA method to predict image quality accurately with great generalization ability. We first use gaussian kernel regression to do the data augmentation owing to the limited dataset. Then, a NR-IQA model based on convolutional neural network (CNN) is trained, named Deep No-Reference Image Quality Assessment (DNRIQA), including five convolutional layers and three pooling layers for feature extraction, and three fully connected layers for regression. Finally, the experiments show that this approach achieves the state-of-the-art performance on LIVE dataset, and further cross dataset experiments prove that the model has excellent generalization ability. Experimental results of this paper is very competitive compared with other algorithms.