No-reference Image Quality Assessment Based on Convolutional Neural Network

Abstract

Image Quality Assessment (IQA) is a classic research topic, whose goal is to design algorithms and give objective values consistent with subjective values of the human visual system (HVS). IQA plays an important role in many image processing applications, such as image enhancement, image compression and reconstruction, watermark addition, etc. In this paper, we present two no-reference image quality assessment (NR-IQA) models based on convolutional neural networks (CNN). One of the biggest challenges in learning NR-IQA model is lack of images with subjective value. Thus, we label images with full-reference (FR) algorithms, the intermediate product “similarity map” is generated when some FR algorithms calculate image quality. We use similarity map or objective score of some FR algorithms to label images separately. The first model uses objective values of SSIM, VIF, GMSD and FSIM respectively, to label images to train the improved VGGNet. The second model uses similarity map generated by FSIM and VSI respectively, to label images to train the improved U-Net. Experiments conducted on the database D215 built in our laboratory show that our second model is comparable to the advanced NR-IQA model.