A Perceptual Distinguish ability Predictor For JND-Noise-Contaminated Images
Abstract: Just noticeable difference (JND) models are widely used for perceptual redundancy estimation in images and videos. A common method for measuring the accuracy of a JND model is to inject random noise in an image based on the JND model, and check whether the JND-noise-contaminated image is perceptually distinguishable from the original image or not. Also, when comparing the accuracy of two different JND models, the model that produces the JND-noise-contaminated image with better quality at the same level of noise energy is the better model. But in both of these cases, a subjective test is necessary, which is very time consuming and costly. In this paper, we present a full-reference metric called PDP (perceptual distinguishability predictor), which can be used to determine whether a given JND-noise-contaminated image is perceptually distinguishable from the reference image. The proposed metric employs the concept of sparse coding, and extracts a feature vector out of a given image pair. The feature vector is then fed to a multilayer neural network for classification. To train the network, we built a public database of 999 natural images with distinguishbility thresholds for four different JND models obtained from an extensive subjective experiment. The results indicated that PDD achieves high classification accuracy of 97.1%. The proposed method can be used to objectivel compare various JND models without
performing any subjective test. It can also be used to obtain proper scaling factors to improve the JND thresholds estimated by an arbitrary JND model. Existing system: Based on the above discussion, we observe that for validating a JND model or comparing it to another JND model, a subjective test is necessary. But performing subjective tests is usually very time consuming and costly, especially for a large number of images. At the same time, the JND thresholds estimated by many existing JND models are not completely accurate, and depending on the image content, a scaling factor. is usually applied on the JND thresholds so that they become more accurate. However, to the best of our knowledge, none of the existing models provides a systematic framework for adaptively setting the scaling factors based on the image content. Fig. 1 illustrates this issue on a sample image for two different JND models. As seen from this figure, when the scaling factor is equal to one, both models produce JND-noise contaminated images that are clearly distinguishable from the original, whereas when using two different scaling factors smaller than one, the noisy images produced by both models look indistinguishable from the original image. Proposed system: The database can also be used to reveal shortcomings of various JND models when applying to images with different content. The database and the code of the proposed metric will be made available at if the paper gets accepted. We present a systematic approach for choosing the scaling factor of an arbitrary JND model based on the image content for a more accurate estimation of JND thresholds. Such an approach can also be used to compare various JND models without doing any subjective test. Note that since conducting subjective tests can be cumbersome, many existing JND models were only tested on a few standard images. However, with the proposed metric, it will be possible to measure the accuracy and efficiency of an arbitrary JND model objectively, and perform statistical significance tests to make statistically valid conclusions. Advantages:
provide a quality score, but the user must interpret the score in terms of distinguishability. Also, it is worth mentioning that there exist some visible difference predictors like VDP, and its variants like HDR-VDP-2 that can be used to predict the visibility of distortions at the pixel level in terms of a probability map. They can also provide an overall image quality score by pooling the computed probability map. But even using such predictors, the user must still interpret the score in terms of the overall distinguishability at the image level. We present a public database of 999 images with image level distinguishability thresholds for four different JND models. We are not aware of any similar existing database. Such a database can be used to generate an infinite number of perceptually-similar or not-similar reference-noisy image pairs. Such image pairs can be used in various applications such as learning-torankbased image quality assessment , perceptual data augmentation for deep learning applications, etc Disadvantages: where _ and _ denote the element-wise multiplication and division, respectively, and c1 = 1, which is used to stabilize the result. Note that due to normalization, the values in the similarity matrix S1 do not change by the magnitude change of elements in F_ R or F_J. To account for this problem. It is known that regions with larger distortions have a higher impact on the human judgment of perceptual quality. In other words, regions with poor quality have a more severe impact on the perceived image quality than the regions with better quality. To consider this issue, we employ a perceptual full-reference image quality assessment metric to estimate the perceptual quality of various regions in J. Modules: Just noticeable difference:
The human visual system (HVS) cannot sense small variations of visual signals whose amplitudes are below the socalled just noticeable difference (JND) threshold due to several physical limitations and spatial/temporal masking . JND estimation is widely used for perceptual redundancy estimationin images/videos for a wide variety of applications such as image/video compression and transmission, image/video quality assessment , watermarking , steganography, etc. In the literature, various models have been developed for JND estimation in static images and videos in both the pixel and subband/frequency domains . For instance, in , Yang et al. proposed a pixel-domain JND model based on the nonlinear additivity model for masking (NAMM), and used it for video coding. In , a JND model was proposed for images based on measuring edge and texture masking . perceptual similarity: The selected patches are then transformed to the perceptual space defined by the learned sparse feature detector (dictionary) through a linear transformation. This way, each patch is represented by a sparse feature vector (code), where the code can be considered as the response of cortical cells in the human visual system. Therefore, by measuring the similarity between the sparse codes of a reference patch and its distorted counterpart, one can measure the perceptual similarity of the distorted patch to the reference patch. For this purpose, we use two similarity metrics to build a single feature vector out of the two input images. The resultant feature vector is then fed to a binary classifier, which is able to determine whether the JND-noise-contaminated image is perceptually distinguishable from the reference image or not. Neural network: Therefore, PDP may have a better ability to deal with JND-noise-contaminated images than SFF or SPARQ. 4) To compute the image quality score, SFF and SPARQ directly measure the similarity between the extracted sparse feature vectors while PDP feeds the extracted sparse features to a neural network for classification. We use a multilayer feed forward neural network as the classifier. To train the network, we use a new dataset obtained through an extensive subjective experiment with 15 subjects on a database of 999 images. In this database, for each reference image, there are four thresholds called
“distinguishability thresholds� related to four different JND models, where the thresholds are obtained as follows. In each trial of this experiment, a reference image was displayed along with a JND-noise-contaminated image produced by one of the four JND models.