No-reference image quality assessment through transfer learning

Transfer learning has emerged from recent years' great success of applying convolutional neural networks to object recognition tasks. In this work, we describe a transfer learning framework that learns an image quality estimator end-to-end in classification or regression. Experiments show this approach achieves state of the art performance on four publicly available image quality databases. Image quality is ignored in visual recognition challenges such as PASCAL VOC and ILSVRC. However, it poses a practical challenge to the design of robust vision applications. This work responds to such challenge without impacting the performance of recognition tasks.