PROJECT TITLE :

A Perceptual Distinguishability Predictor For JND-Noise-Contaminated Images

ABSTRACT:

These models are commonly used to estimate perceptual redundancy in images and videos. You can use a typical JND model evaluation method to test the model's accuracy by using a random noise generator to insert random noise into an image generated by the JND model. At the same time, comparing two distinct JND models, it is preferable to use one that provides an image with greater quality at the same amount of noise energy than the other model. It's time-consuming and expensive to conduct a subjective test in any instance. Full-reference metric termed PDP (perceptual distinguishability predictor) can be used to evaluate whether a particular JND-noise-contaminated image is perceptually discernible from the reference image. For example, the suggested metric leverages the sparse coding idea and extracts an image pair feature vector. After that, a multilayer neural network is used to classify the feature vector. An comprehensive subjective experiment yielded four distinct JND models, which we used to build a public library of photos with distinguishability thresholds. According to the findings, PDD has a classification accuracy rate of 97.1%. An objective comparison of several JND models can be made using the proposed method. It can also be used to improve the JND thresholds estimated by an arbitrary JND model by using correct scaling factors.