Unsupervised Deep Learning for Just Noticeable Difference Estimation

Just noticeable difference (JND) estimates the visual redundancies of the human visual system (HVS), which has been widely applied in perceptual redundancy estimations in images and videos. Existing handcrafted feature based JND models are always inspired by some kind of HVS mechanisms, and have a limited performance for JND threshold estimation. Recently, deep learning has been widely used in various visual tasks and achieved notable success. However, deep learning is difficult to be applied for JND estimation, since it is impossible to build a large-scale dataset with pixel-level-label for training. In this paper, we propose an unsupervised learning based JND model. The underlying idea is to learn the visual redundancy characteristics of HVS by convolutional neural networks (CNN) without labeled data. Specifially, in order to optimize the parameters of the proposed model, three types of prior knowledge, i.e., the image quality, the pattern complexity, and the noise masking ability, are used to estimate the JND map of an image. Experiments demonstrate that the proposed model is highly consistent with the HVS and outperforms the existing JND models.