Evaluation Of Domestic High-Resolution Image Quality Based On GIQE

With the rapid development of high-resolution satellite development in China,the ability of obtaining high-resolution images is increasing.In the face of massive highresolution images,the research on the evaluation method of remote sensing image quality is helpful to screening and classifying the data,which makes it convenient to provide the corresponding data according to the needs of users,and the research will promote the development of the commercialization of remote sensing data services.In the application-oriented remote sensing image quality evaluation method,the interpretability of image is an important index for users to select image and evaluate image availability.With the improvement of spatial resolution of remote sensing image,more types of ground objects appear,and users' demand for image interpretability is also increasing.Therefore,this paper studies the interpretability quality evaluation method for high-quality remote sensing image.The traditional remote sensing image interpretability evaluation is measured by the NIIRS,which is a subjective evaluation method interpreted by experts.,and the GIQE is an evaluation model developed for it.When the relevant parameters of imaging system are known,NIIRS of image can be estimated.There are some problems in using the GIQE model to evaluate the domestic high-resolution images,mainly including: 1.The image reference target used to extract parameters is relatively special(such as target image),which is difficult to find on the general images,but it is greatly affected by the noise when using the edge of natural objects to extract parameters;2.The foreign image platform and quality used to build the GIQE model are different from that of the domestic ones At the same time,the parameters extraction of the model is greatly limited by the actual conditions,and the original model has bias on the prediction results of domestic high-resolution images;3.In the traditional GIQE model,the image interpretability is mainly controlled by the image spatial resolution,and the image interpretability differences are not obvious under the same spatial resolution,and the prediction results of the model sometimes do not conform to the real manual interpretation results;4.GIQE comes from the simulation of artificial interpretation ability and lacks the applicability analysis of machine interpretation ability such as computer classification evaluation of image.In view of the problem that the extraction of edge parameters of natural objects is greatly disturbed by noise,this paper optimizes the acquisition of edge parameters of GIQE model,allows the use of edge images of natural objects to extract parameters and strictly screen them at the same time,mainly judges whether the edge is straight or not,whether the contrast of edge image is significant,and selects the one with straight edge and large gray-scale difference on both sides of the edge as the edge used to extract the parameters.In addition,an adaptive filter is set to filter the edge image according to the distance from pixel to edge to improve the stability of edge parameter extraction.Aiming at the problem of the prediction deviation of the original GIQE model,this paper uses more than 400 GF-1 and GF-2 images to construct the domestic highresolution image evaluation data set,and extracts the GIQE model parameters of each image on the basis of edge optimization and calculates the predicted value of the original model,which is PNIIRS;at the same time,the real interpretation degree ONIIRS value of each image is obtained by manual visual interpretation.Using the least square regression to establish the quantitative relationship between the ONIIRS and PNIIRS of the training set images and obtain the modified model for the original GIQE model prediction value PNIIRS.And this paper uses the verification set images to evaluate the accuracy.The coefficient of determination of GF-1 data modified model is 0.7890;and the coefficient of determination of GF-2 data modified model is 0.8998;the coefficient of determination of GF-1 and GF-2 data modified model is 0.7920.All of the results show that the method of this paper can improve the prediction value of GIQE model and make it closer to the manual interpretation value.In this paper,MTF performance parameters are introduced to improve the GIQE model.Through the study of the relationship between MTF and image quality,MTFNyquist,MTF-50 and MTF-area are selected as the three parameters of image MTF to replace the RER parameters in the GIQE model to build an improved image quality evaluation model.Comparing the influence of different MTF parameters on the fitting accuracy of the model and the accuracy of the validation set,MTF-50 is selected as the MTF performance parameter in the case of neglecting overshoot H and MTF-Area is selected as the MTF performance parameter in the case of considering overshoot H to construct the improved GIQE model,and the expression of the improved model is obtained.Aiming at the problem of computer classifiability evaluated by GIQE,this paper studies the index of image classifiability,selects the TPR(true positive rate)of the target object and the AUC(area under the ROC curve)of the image to quantify the classifiability of the image,analyzes the correlation between the image NIIRS,TPR and AUC,and finds that when it comes to the image of the same area,while NIIRS of image is higher,the TPR and AUC of the image are higher,and the classifiability is better;when NIIRS is at a high level,the classifiability of the image is significantly increased with the increase of NIIRS;but when the NIIRS level of the image is lower,selecting the image with lower NIIRS can reduce the error rate of classification while the image classification accuracy is not significantly reduced.By establishing the qualitative relationship between NIIRS and image classifiability,we can use GIQE to predict the image classifiability in the same area.