Pixel-Level Sonar Image JND Based on Inexact Supervised Learning

The Just Noticeable Difference (JND) model aims to identify perceptual redundancies in images by simulating the perception of the Human Visual System (HVS). Exploring the JND of sonar images is important for the study of their visual properties and related applications. However, there is still room for improvement in performance of existing JND models designed for Natural Scene Images (NSIs), and the characteristics of sonar images are not sufficiently considered by them. On the other hand, there are significant challenges in constructing a densely labeled pixel-level JND dataset. To tackle these issues, we proposed a pixel-level JND model based on inexact supervised learning. A perceptually lossy/lossless predictor was first pre-trained on a coarse-grained picture-level JND dataset. This predictor can guide the unsupervised generator to produce an image that is perceptually lossless compared to the original image. Then we designed a loss function to ensure that the generated image is perceptually lossless and maximally different from the original image. Experimental results show that our model outperforms current models.