Transfer learning for just noticeable difference estimation

The just noticeable difference (JND) measures the visual redundancy of digital images and is widely used in signal processing. Conventional JND models attempt to simulate functional properties of the human visual system (HVS), which are limited by the development of cognitive psychology. In this paper, we propose a novel pixel-wise JND prediction model based on deep transfer learning. Since it is almost impossible to manually label each pixel's visibility threshold, lacking labeled training data is the crucial issue. Transfer learning addresses the problem of insufficient training data. We found an underlying correspondence between full reference image quality assessment (FR-IQA) and JND estimation, which implies that knowledge related to FR-IQA can be applied to JND estimation. To quantify the intrinsic association between JND estimation and FR-IQA, a local perceived discrepancy (LPD) index is deduced. With the guidance of the LPD index, a JND predictor based on residual dense network (RDN) is designed to discover good representations of visibility limitation from annotated image quality databases. Subjective viewing test experiments show that our model outperforms the state-of-the-art JND models. Furthermore, we apply our model to image compression, and around 14.42% of the bit rate can be reduced by removing visual redundancy.