Research On Objective Evaluation Metrics For Image Enhancement

The image enhancement algorithm aims to enhance part of the region of interest of the image in a targeted manner for different application scenarios,thereby improving its visual effect and perceived quality.Most of the performance evaluation indicators of enhancement algorithms at this stage are designed based on the pixel errors of the images before and after enhancement.However,such quality standards based on pixel errors only rely on the average difference between the pixel values of the image before and after enhancement to measure the quality of the enhancement effect.Therefore,it is impossible to describe various types of local specific distortions and changes in image perception quality produced by different enhancement algorithms during the enhancement process.Its performance measurement results for different enhancement algorithms often do not conform to human subjective judgment.In view of the above problems,this paper starts from two enhancement algorithms,multi-exposure image fusion and super-resolution reconstruction,explores its enhancement methods,enhancement purposes,and specific distortions that are prone to occur,and builds a reasonable quality measurement model.The specific research results are as follows: First,a quality assessment algorithm for image super-resolution with distortions based on wide-range distortion dimension integration was proposed to solve the problem that various types of specific distortions generated by current super-resolution reconstruction algorithms cannot be effectively measured.The algorithm performs multi-scale scaling on the input image,and extracts distortion feature descriptors of multiple perception domains such as color domain and texture domain at different scales.Then,the corresponding coefficient regression models are established for the same features at different scales and different features at the same scale to generate distortion ratio coefficients in each perception domain.Finally,the algorithm constructs a two-stage integration strategy,fuse the obtained ratio coefficients to obtain the widerange distortion quality description features of the enhanced image,and map them to quality prediction scores.Experimental results show that the algorithm has the ability to accurately describe a variety of specific distortions generated during the reconstruction process,and the evaluation results of the enhancement effects of different super-resolution algorithms are also highly consistent with human subjective perception.Second,a super-resolution reconstruction algorithm based on perceptual adaptation was proposed in order to solve the problems that optimization functions of super-resolution reconstruction algorithms cannot accurately reflect the true perceived quality of the generated images.In this study,the super-resolution reconstruction network was built and pre-trained by stacking of residual feedback modules,the pixel error between the low-resolution images and the high-resolution images was then used as the loss function.Next,the quality-based perceptual network was built by multiple sets of perceptual features and pre-trained using the super-resolution evaluation datasets.The low-level texture features and high-level semantic features were extracted from the images using quality-based perceptual network,and then feedback differences of the features to the reconstruction network,which used for the parameter optimization.The superresolution reconstruction of low-resolution images was then performed using the trained super-resolution network.The experimental results revealed that the generated super-resolution images by the proposed approach have more realistic texture details and sharper edge contours as well as better visual experience,which is significantly better than the existing super-resolution algorithms.Third,an image quality assessment algorithm based on luminance distribution combined aesthetic perception distribution for multi-exposure image fusion was proposed.It is difficult to evaluate the perceptual quality with currently general quality assessments because image information such as illumination distribution and aesthetic distribution are easily destroyed during the fusion process,which seriously affects the illumination consistency and aesthetic characteristics of the fused images.To solve the above problems,the luminance error regressor was constructed using multi-channel twin network and pre-trained using the image dataset that obtained from relative ordering between the label values of subjective scores in the quality evaluation database.The luminance distributions of the fused images were then obtained from the pretrained luminance error regressor.Next,the multi-tasking aesthetic evaluation was pretrained on a large aesthetic database,which used for extracting aesthetic distribution features and aesthetic targets in the fused images.A shallow comprehensive network was used to learn the mapping function between the fused attributes and the label values of subjective scores.The experimental results showed that the proposed approach could reliably quantify the luminance distribution and aesthetic perception distribution and accurately predict the perceived quality of the fused image,and has a good consistent with performance and subjective judgments between different fusion algorithms.