Research On Remote Sensing Image Quality Evaluation Based On Convolutional Neural Network

During the whole life cycle of remote sensing images,they will undergo imaging,compression,transmission,storage and other processes.Each process will bring different types and degrees of distortion,which will affect the analysis and understanding of images and other subsequent image processing.At present,most of the remote sensing image quality assessment metrics are based on simulated images,which are greatly different from the distortion of authentic images.Moreover,there is no unified evaluation standard.these problems limit the development of authentic remote sensing image processing.Therefore,establishing authentic remote sensing image databases and studying the objective quality assessment algorithm are urgent problems to be solved.Convolutional neural network processes target task by simulating learning process of human brain,and its greatest advantage is the automatic learning of features.Authentic remote sensing images have aliasing distortion types and complex textures,making it difficult to extract effective quality features by humans.Therefore,this paper has carried out the research on how to apply convolutional neural networks to remote sensing image quality assessment.According to the practical application scenarios of remote sensing image quality assessment,we construct a database based on comparative quality.And a CPNet(compare-net)structure based on image quality difference is proposed for image comparative quality evaluation.At the same time,we study the remote sensing image with non-uniform distortion,and propose a usability assessment algorithm based on convolutional neural network with remote sensing image characteristics.The main research contents are as follows:(1)A convolutional neural network CPNet is proposed for image comparative quality assessment.Firstly,by analyzing the influence of convolutional neural network structure related parameters on performance,a reasonable network is constructed.Secondly,the two-image input and feature subtraction in the network are used to obtain the difference characteristics of the quality between two images.And then combined with the label of comparative quality of the image pair to complete the classification learning.The experiment of image pairs with the same reference image and the different reference image in the LIVE database proves that the accuracy of the proposed algorithm is better than other algorithms,and the performance of the algorithm is more stable.(2)An authentic remote sensing image database of environmental satellites based on comparative quality is constructed and CPNet is applied to the comparative quality assessment of remote sensing images.In the images from HJ-1A/1B satellites,select 70 representative locations to ensure that the feature types are comprehensive.Then select 5 images at different times for each location,and following are the illumination and geometric consistency corrections.Therefore,the constructed database contains 350 remote sensing images with the size of 600×600.The comparative quality scores of the five images in each group were obtained after subjective assessment and analysis of all subjective scores.Finally,the experimental performance of CPNet on the database of this paper verifies the robustness of CPNet and the rationality of the database.(3)A remote sensing image usability assessment algorithm based on ResNet with Sobel edge and LBP texture feature map is proposed.The algorithm combines the characteristics of human visual system to solve the quality assessment for high resolution remote sensing images with non-uniform distortion.Considering that the integrity and clarity of the contours and internal details are important factors in the subjective local quality assessment,this paper adds Sobel edge feature map and LBP texture feature map as a prior knowledge to the training process of ResNet for local quality assessment.Finally,the image usability score is obtained by the predicted local quality score and the weight map.The experimental results on the remote sensing image database based on usability demonstrate that the edge and texture feature map can effectively improve the accuracy of the algorithm.