Image Quality Evaluation Based On Visual Perception And Image Features

Image quality evaluation is one of the basic research contents in the field of image processing.In recent years,with the continuous progress of communication technology and the rapid popularization of intelligent terminal equipment,the application fields and application requirements of digital images are growing rapidly.Digital image is widely used in the fields of information communication,biomedicine,industrial production,geological exploration,meteorological forecast,military defense and space exploration.In the application process,the image quality has an absolute impact on the application value,highquality images are more conducive to people's reception,understanding and processing of image information.However,in practical application,due to the limitation of the actual environment and technology,the inevitable noise will be introduced in multiple stages of image acquisition,transmission and storage,resulting in image distortion and quality damage,thus affecting the effect of practical application.In order to accurately and effectively measure the change of image quality,the researchers began to explore effective methods for evaluating the quality of images.The evaluation of image quality can be explored from subjective and objective perspectives.In the objective evaluation,the FR-IQA can obtain a complete reference image and obtain better evaluation performance.In FR-IQA research,the method combined with visual perception characteristics can often achieve more consistent evaluation performance with subjective evaluation.However,most combination strategies are relatively simple and lack strong theoretical support.By analyzing the calculation process and design ideas of classical methods,this paper fuses visual perception characteristics and image characteristics based on machine learning and convolutional neural network respectively,and proposes two new image quality evaluation methods: RF-IQA and FPN-IQA.RF-IQA extracts the phase consistency feature,edge feature,color feature and texture feature of the image,then constructs the feature similarity vector of 11 dimensions,and finally adopts the method of machine learning to learn and train the image quality regression model.FPN-IQA introduces the deep learning method to extract the image features with the fusion of visual perception characteristics,FPN based on Resnet-50 network is used to extract the image features,then the image quality is evaluated by calculating the feature similarity.Experiments in several databases show that the RF-IQA and FPN-IQA have better overall performance and the evaluation performance of different distortion types has better robustness.