Research On Image Bit-Depth Enhancement Algorithm And Quality Assessment Based On Semantic Prior

High bit-depth images can provide richer colors and more detailed content with higher visual quality.With the rapid development of high dynamic range display technology and the continuous improvement of human living standards,demand of people for high bit-depth image is also increasing day by day.Since the bit-depth enhancement algorithm can realize the conversion from low bit-depth images to high bit depth ones,it has been extensively studied in recent years.However,most of the existing algorithms ignore the semantic information of the image,the reconstructed image often has the false contours effect on flat area and the detail of textures or edges is lost.Moreover,there is currently a lack of dedicated quality evaluation algorithms for bit-depth enhanced images,it is difficult to evaluate the quality of the reconstructed images accurately and reliably.In order to solve the above problems,this paper comprehensively considers the characteristics of bit-depth enhancement task and human visual perception,and conducts research on image bit-depth enhancement algorithm and bit-depth enhancement image quality assessment algorithm based on deep learning.The main research contents and innovation results are as follows:1)This paper introduces the semantic category prior and the semantic structure prior to the bit-depth enhancement task for the first time,and proposes an image bitdepth enhancement algorithm based on the semantic guided generative adversarial network.In order to generate more realistic and natural high bit-depth images,semantic category priors are introduced into both the generator and discriminator.The semantic guided residual block of the generator generates modulation parameters based on the semantic segmentation map,and modulates the intermediate features of the input image to generate texture details consistent with the inherent content of the image.The discriminator takes the cascade of the image and its semantic segmentation map as input,and uses auxiliary classification branch to identify the semantic category of the image to enhance the ability of the discriminator in identifying true and false images.In addition,in order to further suppress the false contour on flat area,a gradient loss based on the semantic structure prior is proposed.The flat area of the image is given a higher loss weight than the non-flat area.Experiments show that the proposed algorithm can restore finer detail of texture and edge while suppressing false contours on flat area.The subjective and objective evaluation results on multiple datasets exceed other traditional and deep learning methods.2)This paper proposes the first non-reference image quality assessment algorithm for the distortion of bit-depth enhancement.In order to effectively extract the features of common false contours and over-blurred distortion of bit-depth enhanced images,a two-branch network with image branch and gradient branch is proposed.The gradient branch takes the gradient map as input to strengthen the extraction of structural distortion features.In addition,in order to make the quality score output by the network more consistent with the subjective quality score,a semantic attention module is constructed to simulate the characteristics of human visual perception.This module uses the image semantic segmentation map as prior condition to give different attention to different content of the image.Moreover,in order to make the quality assessment model converge better,a pre-trained ranker is proposed to calculate the ranking loss,which indirectly optimizes the ranking index.Experiments show that the proposed algorithm outperforms the other state-of-the-art algorithms on the bit-depth enhanced image quality assessment dataset.