Image Quality Assessment And Its Application In Image Denoising

With the advent of the multimedia era,image and video have gradually developed into the most commonly used information carriers in human activities,and become the main way for people to obtain original information from outside.Since the quality of image affects people's subjective feelings and information acquisition directly.There-fore,objective and effective assessment of image quality has become an indispensable means in the field of image information processing.By studying the human visual perception characteristics,this dissertation deeply goes into the human visual characteristics and the physical mechanism of the image,and focuses on the objective quality assessment and application of natural images.The main work includes the following three aspects:(1)The full-reference image quality assessment mechanism is studied,and the lo-calized feature of the image is weighted by using the pooling function modeled by human visual characteristics.Aiming at the visual saliency characteristics in the human visual attention mechanism,a full-reference image quality assessment algorithm is proposed.The algorithm defines the edge structure information of the image as the combination of horizontal gradient,vertical gradient,and gradient direction,and then uses the human visual property model to weight and sum the local quality map to obtain the final image quality score.When calculating the gradient direction similarity,the algorithm makes a new definition,and the feasibility and effectiveness of this definition are verified by experiments.Aiming at the sensitivity of human visual contrast,a full-reference image quality assessment algorithm based on feature similarity is proposed.The algorithm uses the monogenic signal theory and the log-Gabor filter to combine the sensitivity of the human visual contrast with the similarity of the gradient magnitude to obtain the local quality map.Then,by using the monogenic phase coherence to construct the pooling function,the final image quality assessment score is obtained.The algorithm applies local features in Riesz's analytical space,including local amplitude,local phase,and local direction,to image quality assessment,and achieves good results.(2)The no-reference image quality assessment are modeled in both dedicated and general purpose.Aiming at the problem of noise image quality assessment,a special type of non-reference image quality assessment algorithm is proposed.The algorithm combines the human visual contrast sensitivity characteristics with the bi-dimensional empirical mode decomposition,and proposes a frequency mapping of visual contrast sensitivity,and obtains good noise distortion image assessment performance.Especially in the case of high image distortion,some classical algorithm assessment scores are no longer related to subjective assessment scores,our algorithm can still maintain good assessment accuracy.In addition,the algorithm does not require training,making the assessment performance of the algorithm universal.Aiming at the general-purpose no-reference image quality assessment,a quality assessment algorithm based on image entropy is proposed.The algorithm introduces the concept of two-dimensional entropy into image quality assessment,and combines log-Gabor filtering and visual saliency detection to extract color image features,and then obtains the final image quality assessment score by means of support vector machine.Experiments show that features such as two-dimensional entropy accurately reflect the structural information of the image,and have achieved good results,reflecting the excellent performance and good generalization ability of the algorithm.(3)The image processing technology based on image quality assessment model is investigated.Aiming at the specific applications such as image denoising,a deep learning denoising algorithm based on image quality assessment guidance is proposed.The algorithm organically combines the pixel loss function and the perceptual loss function,and uses the image quality assessment network to guide the image denoising network training,and achieves good denoising effect.The denoising network adopts the residual structure to convert the direct training restoration image into training an additive noise filter,which reduces the training difficulty and accelerates the convergence speed of the network parameters.The quality assessment algorithms proposed in this dissertation have carried out rich experiments on the published quality assessment database,and obtained the results of performance indicators and competition rankings,indicating the subjective and ob-jective consistency of the algorithm,reflecting the good performance of the algorithm.Finally,the denoising experiment was carried out on the widely used image denoising database and quality assessment database,and the higher objective assessment scores were obtained.At the same time,the restored images more in line with human percep-tion were obtained.This dissertation further broadens the ideas for the basic research and extensive application of image quality assessment,and also provides a reference for future development.