Towards On-board Image Quality Assessment For Remote-Sensed Images

With the advanced development of remote sensing(RS)techniques,the RS image processing has entered a big data era,with image data growing exponentially.However,influenced by ato mo sphere motion,atomosphere absorption and dispersion,platform motion,and electronic component damages of sensors,captures images are often distorted by cloud-cover and other image-content distortions,which might generate invalid data.Most of the on-board systems of remote sensing satellites are unable to detect and eliminate invalid data effectively,which results in big waste of the computing and storage resources on-board,and the space downlink bandwidth.Here,being concerned with the automatic detection,recognition and degree evaluation of remote sensing invalid data,in this thesis,we mainly concentrated on the cloud-cover and the other common image distortions,and then proposed an effective cloud-cover assessment metric together with a series of quality assessment(QA)metrics for remote-sensed images.The main contributions of this thesis can be summarized as follows:(1)In terms of the detection and evaluation of the invalid data caused by a large amount of cloud-cover,the exisiting cloud-cover assessment methods cannot make a good balance between performance and computational complexity.Unlike them,we proposed a both effective and efficient cloud-cover assessment approach based on the natural scene statistics(NSS)of panchromatic RS images.In this thesis,we found that the distributions of the mean subtracted contrast normalization(MSCN)coefficients of panchromatic cloudy images and cloud-free images are different in shape,and the asymmetric generalized Gaussian distribution(AGGD)can well model those differences embedded in shapes.The resultant AGGD paramters are futher fitted by a multivariate Gaussian model(MVG),providing a rich feature representation.The amount of cloud-cover of one test image is thus defined as the distance between the MVG fatures of cloud-free images and those of cloudy images.Experimental results testing on the Landsat 8 panchromatic cloud image dataset demonstrate that the proposed NSS-based cloud-cover assessment method outperforms the Landsat 8 metric.Moreover,it has low computational complexity.(2)Regarding the QA of the RS invalid images induced by serious image distortions,we first proposed a full-reference(FR)RS image QA method by separating image into structure and detail portions with singular value decompostion(SVD),considering the situations where there exist reference images,such as the procedure of image compression.Different types of distortion function differently on different parts of images.For example,noise mainly distorted image details and blur generally smooths edges and structures.Therefore,these two portions should be evaluated separately by different methods.In this thesis,image gradient and contrast similarities are adopted for the QA of structure portion,and the normalized peak signal-to-noise ratio(PSNR)is used for the QA of detail portion.The final image quality score is computed as the weighted multiplication of the QA results of the two porpotions.Experimental results in the RS image QA database and public visible image QA databases show that the proposed metric has better performance than those state-of-the-art FR methods.Although the above structure-detail decomposition QA method has addressed the issue of the different influences of different types of distortion,the extracted features and corresponding feature regression method seem to be not very consistent with the ultimate perceptual results of human visual system(HVS).Non-negative matrix factorization(NMF)has been proved to be able to learn the part-based features from natural images.Inspired by that property,in this thesis,we proposed a perceptual FR QA metric for RS images using NMF for extracting features from reference and distorted images.The similarities between the extracted features are adopted to measure the quality degradations.Finally,extreme learning machine(ELM)is used to map the similarities into image quality.We tested the NMF-based method in both the RS image QA database and visible image QA databases.Experimental results demonstrate that the proposed NMF-based perceptual FR QA method outperforms other FR methods with higher quality prediction accuracy.(3)With respects to situations where there exist no reference images on-board,for noise distortion,we developed a no-reference(NR)QA metric for RS images based on subband kurtosis.Noise is one of the most common image distortions that degrade the interpretability of RS imagery and is produced by such factors as atmospheric scattering,photon noise,shot noise and electronic failures.Noise generated by the aforementioned factors on an onboard optical RS system is usually data independent and additive in nature.We analyzed the properties of clean and noisy RS images,and found that the distributions of their Discrete Wavelet Transform(DWT)subband coefficients are quite different in shape.Clean(noise-free)images exhibit leptokurtic,peaky distributions with heavy tails,while those of noisy RS images tend to be platykurtic,with less peaky distributions and shallow tails.Such a property is well modeled by the sample kurtosis.Experimental results in the additive Gaussian noisy RS image database and the public visible noisy image databases show that the proposed kurtosis-based method has better assessment performance than other noise-specific NR methods as well as general-purpose NR methods.(4)Focusing on the on-board QA for blur RS images,inspired by the property that the statistics of image gradient magnitude(GM)follows Weibull distribution,we proposed a content-robust NR metric via parameterizing the GM the Weibull distribution.We also adopt skewness to measure the asymmetry of the GM distribution.In order to reduce the influence of image content and achieve more robust performance,divisive normalization is then incorporated to moderate the extracted Weibull features and skewness.The final image quality is predicted using a sparse extreme learning machine which is able to deliver more robust and generalized performance.Performances evaluated on the RS blur image dataset and the public visible blur image datasets demonstrate that the proposed method is highly robust with image content variations,and outperforms other NR blur QA metrics.In addition,our method has low computational complexity.