Deep Learning Based Just Noticeable Difference Modeling Research

Due to the limitation of the human visual system(HVS),it is difficult to detect redundant information in the image or video.The redundant information can only be detected when the amplitude of pixel change exceeds the threshold of the HVS.The minimum amount of visual content change perceived by the HVS can be represented by the just noticeable difference(JND).JND modeling is a hot spot and a challenging research topic in the field of visual perception.Although the traditional JND model inspired by the working mechanism of the HVS has achieved some good results,the current research on the HVS is still limited,which limits the development of this kind of JND model.In recent years,deep learning technology has achieved significant success in the field of image processing,and some JND models based on deep learning have shown excellent performance.However,the existing JND model based on deep learning still has a lot of room for improvement.This thesis will deeply study the JND model based on deep learning technology.The specific contents are as follows.(1)This thesis studies the RGB full-channel JND modeling for the first time and proposes the RGB-JND-NET model.This model is based on deep learning technology and fully considers the stimulation of full-color channels,and makes a preliminary exploration of accurate JND modeling of full-color space.Secondly,to effectively monitor the generation of JND,this thesis proposes an adaptive image quality assessment(A-IQA)network module.Finally,considering the relationship between the human visual attention mechanism and JND,this study put the visual attention feature of the image into the network,so that the network can improve the accuracy of JND modeling by the relationship between human visual attention and JND.The experimental results show that the performance of the RGB-JND-NET model is better than that of the relevant JND models.(2)This thesis proposes JND in Arbitrary Color Space(JND-ACS),which is a full-channel JND model based on deep learning.It can apply to any color space,such as HLS,HSV,XYZ,LAB,LUV color space,etc.JND-ACS considers the full-channel characteristics information of a specific color space,and it is based on unsupervised end-to-end learning to optimize the generation of JND.The experimental results show that JND-ACS can achieve accurate modeling of JND in a specific color space.(3)This thesis proposes the ACo L-JND model based on Adversarial Complementary Learning(ACo L)technology,which is mainly composed of two parallel convolution neural networks and a dynamic erasure mechanism.This design forces the corresponding network to extract new complementary features by erasing the features extracted from one of the networks.Furthermore,the model also introduces pattern masking and contrast masking as prior knowledge into the network to improve the accuracy of JND estimation.The experimental results show that the accuracy of this model to estimate JND is higher than the existing JND models.