Research On Image Quality Assessment And Enhancement Algorithms With Applications

Image is an important source for human to get information.It is one of the ways for people to share their lives and express their emotions.It is also an important data type for computer vision tasks.Image quality varies due to problems with shooting equipment,image transmission,etc.Low-quality images seriously affect people’s visual needs and the accuracy of subsequent vision tasks.Objective and accurate image quality assessment methods can select low quality images for image quality enhancement algorithms and also provide optimization feedback.Currently,the small amount of data and the difficulty in obtaining reference images are the difficulties in the image quality assessment research area.Image superresolution and image denoising as non-single-solution image quality enhancement tasks still have more room for improvement.Therefore,this paper is focused on image quality assessment and image quality enhancement algorithms and applications.At the algorithm level,this paper focuses on the difficulties of image quality assessment,image super-resolution,and image denoising.At the application level,this paper applies image quality assessment methods to detect camera lens local resolution failure,and applied image denoising algorithm to denoise the water surface reflection image in swimming pool dangerous behavior detection,and finally proposes a natural bokeh effect rendering method to improve image aesthetic quality.The main contributions of this paper are as follows.(1)In the image quality assessment research,for the problems of difficult to obtain reference images and small amount of image data,this paper proposes a biovision-inspired image quality evaluation model with multi-scale image input,multi-scale network structure,multi-scale output features and no reference image needed at all.Inspired by biological vision,a module is designed to simulate the multi-scale contour response of the human eye optic nerve to images at different distances,so that the model focuses more on image edge contour sharpness rather than image color features.Inspired by the retinal ganglion cell perceptual field mechanism,a central attention peripheral suppression module was designed,while combining luminance information with different quality weights.The model uses a two-stage approach to solve the problem of small amount of data.The model was successfully applied in a multi-task study of image distortion type classification and quality assessment after adding a classification task branch.(2)For the image super resolution with large factor,a model based on image local feature association is proposed.At present,many super resolution models are mostly applied below 4x.In order to address the problem of large factor,the local feature association module is proposed to establish association degrees for image blocks in different regions,so that image blocks with similar features can draw on each other.Super resolution images are generated by pre-trained jump-connected generative networks,and then fuse the images of different resolutions by multiresolution step-by-step enhancement module.The features of different resolution images are fully fused to further optimize the details of textures,contours,colors and other features of super resolution images.(3)For the problem that it is difficult to obtain real noise-clean image pairs,an image denoising model based on noise-space feature representation is proposed.Most of the current supervised deep learning denoising models use artificially added Gaussian noise to generate noisy images.The difference distribution between the artificially synthesized noise and the real noise leads to the poor performance of the model on real noisy image denoising.To address the above problems,this paper proposes a self-supervised contrast learning based approach to learn noisy features without noise labels by two positive sample generation methods,image panning and noise self-similarity search.Meanwhile,we applied the denoising model to the task of dangerous behavior detection in swimming pools,and successfully solved the problem that the water surface image reflection noise affects the subsequent detection.It fully illustrates that image quality enhancement tasks such as image denoising are important as pre-processing functions for subsequent vision tasks.(4)For the limitation of camera lens resolution detection location,a method for local resolution failure detection of camera lens based on the idea of image quality assessment is proposes.Most camera lens manufacturers currently detect the top left,bottom left,top right,bottom right and middle areas of the lens.In order to detect the resolution failure at any position of the lens,this paper indirectly reflects the resolution of the lens through the captured image based on the principle of image quality assessment,and proposes a calculation method based on luminance and gradient as the index to measure the resolution of the lens.The feasibility of the proposed method is illustrated by correlation experiments and visualization analysis.(5)In order to improve the image aesthetic quality,this paper proposes a refocusing image generation method based on depth estimation of selectable focal plane based on the characteristics of natural bokeh effect.Firstly,the depth information of objects in the image is obtained through the depth estimation module,and then different focal planes are selected through the background subdivision module,and different blur radii are calculated to make the bokeh rendering effect more diverse.The visualization results show that the method satisfies the bokeh effect characteristics of clear objects in the focal plane and blurred objects out of focus.This paper solves some of the difficulties of the image quality assessment and enhancement,and also successfully applies the image quality assessment and enhancement algorithm in the fields of camera lens resolution detection,image preprocessing,and image aesthetic enhancement,which fully illustrates the significance and importance of the research in this paper.