Research On Attitude Jitter Estimation And Image Recovery For Satellite Pushbroom Camera

As for the pushbroom remote sensing images,the satellite attitude jitter is caused by the light pressure or the vibration of the flexible device during the imaging process of the onboard camera,which affects the imaging process of remote sensing.In this paper,the topic of the swept remote sensing image distortion caused by the attitude jitter is discussed,and the followings are the main contents:Aiming at the characteristics of pushbroom satellite remote sensing,the remote sensing image was preprocessed by deep learning theory.Based on the principle of deep learning,a criterion for image segmentation is proposed.The theoretical analysis of deep learning algorithm is given.The principles of deep learning in image segmentation task are introduced.A deep learning cloud-based algorithm based on remote sensing image is designed.Aiming at the lack of computing power in the on-orbit satellite platform,a lightweight image de-cloud model combining discrete wavelet transform algorithm and cloud layer detection algorithm is introduced to achieve the pixelwise cloud detection on satellite platform.The image cloud amount is quantitatively measured,and the image samples are filtered and selected for the further jitter estimation tasks.Considering the imaging parallax system between different spectral bands of multispectral remote sensing camera,mathematical modeling and analysis of multi-spectral image distortion is given.Then a new image registration method is proposed to obtain images pixel offset between different spectral bands.In view of the shortcomings of existing registration algorithms,this paper proposes an improved L-K algorithm to obtain pixel offset at the sub-pixel level.According to the transformation between the image coordinate system and the star body coordinate system,the relative attitude jitter under the body coordinate system is obtained.A time-domain and frequency-domain based deconvolution algorithm is introduced to convert the relative jitter into an absolute coordinate system in the time coordinate system.In this paper,a deconvolution algorithm combining time-frequency domain is proposed,which further improves the accuracy and stability of the deconvolution algorithm.Aiming at the ill-posed characteristics of the deconvolution algorithm,a pose smoothing regular term and a polynomial-based regularoptimization term are introduced to obtain a stable solution of deconvolution and improves the efficiency of the algorithm.A kalman filter is proposed to fuse the jitter information with oboard gyros and star trackers to calibrate the onboard attitude sensors.In order to detect the distortion of single grayscale remote sensing image,this paper introduces a deep learning algorithm to recover the grayscale distortion image and detect the jitter curve.In order to train the deep learning model,firstly,the multi-harmonic sinusoidal distortion curve is simulated according to the satellite distortion dynamics.And the distortion-clear image pair is obtained by the interpolation algorithm.Considering the learning ability of image convolution operations,a multi-layer convolutional neural network is built.A residual convolution block is introduced to fit the feature difference of the distortion-clear image pair.Besides,the extended Sigmoid activation function is introduced to increase the nonlinear fitting ability of the network.After obtaining the distortion curve,the image interpolation model is built to achieve the end-to-end network learning model.Aiming at the training of network parameters,a combined loss functions are designed,and the Adam optimization function is introduced to train the network parameters.An image augmentation algorithm is used to reduce the overfitting of the model.In order to further improve the accuracy of the network model,an adversial network based on gradient penalty term is introduced as a loss function to guide the network to optimize more efficiently.The simulation analysis proves that the introduced loss function can effectively recover the distorted image and detect the distortion curve.For the reconstructed images,the image indexing algorithm and the image segmentation algorithm are introduced to test the image distortion-recovery image pair.The experimental results verify the performance improvement after the deformed image correction.Aiming at the distortion information obtained by multi-spectral image and single-band remote sensing image,this paper proposes a Kalman filtering algorithm to fuse multisource distortion information to further improve the accuracy of distortion information and improve the image restoration quality.Aiming at the problem of large error of the jitter estimation algorithms under extreme features,a blurred star point jitter estimation algorithm based on a fully connected neural network is introduced to further improve the robustness of jitter estimation and deformed image restoration.