1. Introduction
A star sensor is an optical device that measures the positions of stars to determine attitude or orientation. This kind of instrument has been widely used in space missions and deep space exploration, allowing high precision astronomical navigation. A star sensor works by first imaging the starry sky, and then determining the star centroid coordinate information of stars using star extraction and location methods. Once this information has been obtained, the star positions can be compared with the known absolute positions from a star catalog. Finally, based on this comparison, the three-axis attitude of the star sensor relative to the inertial coordinate system can be calculated. Generally, the star sensor works under static condition and it is typically used during the steady flight state of the carrier. It is assumed that the navigation star and the star sensor are relatively stationary during the exposure time, and the star point is imaged at a fixed position on the image plane. When the star sensor is implemented under dynamic conditions, during exposure time, the star forms a trajectory image on the image plane, leading to the degradation of the signal-to-noise ratio (SNR) and decreased accuracy of the star centroid, which may even prevent extraction of the star position information, thus affecting the overall attitude accuracy of the star sensor.
Therefore, improving dynamic performance is needed to improve the performance of the star sensor. The key focus of efforts to improve the dynamic performance of the star sensor is to optimize the denoising and restoration algorithms for blurred star images.
Many scholars have proposed denoising and restoration algorithms of blurred star images. Bezooijen et al. [
1] proposed a time-delayed integration (TDI) method. In this method, the motion blur of the star was reduced using a special hardware sequential circuit, which improved the SNR of image. However, this method only eliminated motion blur in the y-direction, so an image-processing algorithm is required to further improve the SNR. Based on the TDI method, Pasetti et al. [
2] compensated for the effect of motion by oversampling and binning. Sun et al. [
3] established a motion model of the star centroid by using the angular velocity provided by a gyro, then the degradation function of the blurred star image was obtained and the Richardson Lucy (RL) algorithm [
4,
5] was used to restore the blurred star image. However, the angular velocity provided by the gyro drifts with time, which can also seriously affect the accuracy of the determination of the star centroid. To address this, Sun et al. [
6] revised the gyro data by applying an Extended Kalman filter to improve the accuracy of the motion model. Ma et al. [
7] proposed a multi-seed-region growing technique to preprocess the star image before restoration, but the size of the filtering template was limited, which caused a loss of star energy. Zhang et al. [
8] denoised the blurred star image by using adaptive wavelet thresholding, and then restored the star image with an improved Wiener filter. Sun et al. [
9] removed background noise by correlation filtering and morphological filtering, then determined the angle and length of the blurred star by using an image differential method and the star centroid was calculated. However, these methods obtained the star centroid from a blurred star image, which limited the accuracy.
Given the shortcomings of present methods for the denoising and restoration of a blurred star image, we propose a restoration method for a blurred star image under dynamic conditions. In this approach, the kinematic model of the star centroid and the degradation function of a blurred star image under different conditions are first analyzed. Then, an improved curvature filtering method based on energy function is utilized to remove the noise and improve the SNR of the star image. Finally, the termination condition of the iterative equation in the RL algorithm is established by using the star centroid coordinates obtained from three consecutive frames of restored images, allowing good restoration effect of the blurred star image and high accuracy of the star centroid determination.
3. Blurred Star Image Denoising
The original star image contains a significant amount of additive noise, which includes salt and pepper noise, Gauss white noise, and Poisson noise [
11]. Additive noise is unrelated to the original information of the image, but destroys the image signal by superposition. When a star sensor works for a long period of time, there will be electronic thermal noise and high frequationuency electromagnetic interference noise in the image. These noises have the characteristics of high energy and random distribution, and are difficult to estimate. In this study, the Gauss model is used to construct an original star image, based on which a blurred star image is obtained under the dynamic condition of
according to the parameters of
Table 1, and Gauss noise with mean value of 0 and variance of 30, 50, and 70 are added to it, respectively. Then the traditional RL algorithm is used to restore the blurred star image with different noise for 100 times, and then the star centroid coordinates in the star images can be calculated. Finally, the average value of every 10 star centroid coordinates is obtained, and the star centroid error can be obtained through the comparison with the true star centroid coordinate. The error curves (considering the X axis as an example) are shown in
Figure 4.
Figure 4c–e shows that the larger the noise is, the worse the restoration effect is, because when noise exists in the star image, the RL algorithm will amplify the noise. Besides, once the noise is too large, the star will be submerged, resulting in a significant decrease in the accuracy of the star centroid. Therefore, denoising of the blurred star image is requationuired to achieve a highly accurate star centroid location.
represents the spatial coordinate of the image. We use
to represent the current reconstructed image, and
,
,
,
are the partial derivatives, so the image curvature can be described as:
To avoid the complexity of the explicit calculation of Gauss curvature, Gong [
12] assumed that the surface of the original image was piecewise developable, which meant the Gauss curvature was zero everywhere, allowing a good image smoothing effect with edge protection. To meet the assumption that the image is piecewise developable, the gray value of each pixel is directly adjusted to the tangent plane of the neighboring pixels, and then the curvature information of the image is implicitly used to protect the details of the image. The method uses the principle of minimum distance adjustment, which corrects the gray value using the surface closest to the current pixel among all the tangent planes composed of neighboring pixels.
The Gauss curvature filter has a good effect in dealing with Gauss, salt and pepper noise. The Gauss curvature filter requationuires several iterations to eliminate the noise and single iteration is generally insufficient. However, multiple iterations may smooth the star excessively, which will affect the energy distribution of the star and have a great impact on the image restoration. Therefore, an improved Gauss curvature filtering method based on energy function is proposed in this paper to denoise the image. In the algorithm, the distance between the pixel
and the tangent plane formed by the neighboring pixels in the 3 × 3 window should be calculated. As shown in
Figure 5, the neighborhoods are divided into four kinds of diagonal tangent planes and four kinds of minimum triangular tangent planes, and the combination of diagonal tangent planes and minimum triangular tangent planes are used as basic projection operators, the advantages of which are as follows. First, the diagonal tangent plane enhances the connection of neighboring pixels and can effectively suppress the Gauss speckle noise in star images. Second, the minimum triangular tangent plane has a better effect on the salt and pepper noise.
In Equation (17),
denotes the gray value at
,
represents the distance between
and the tangent plane of its neighboring point, then the gray value of the pixel is corrected using the minimum distance adjustment method.
In Equation (18), represents the corrected gray value at . By using a sliding window to filter each pixel of the image, the Gauss curvature filter of the image can be completed.
To solve the uneven energy distribution of the star caused by the multiple iterations of the curvature filter, a new energy model is established based on the energy minimization theory to distinguish the star and noise. Then, the improved curvature filter is used to obtain the final estimated gray value of the star and the energy of the star contaminated by the noise can be estimated and restored.
In this paper, the energy function of the pixel is defined as
, which can represent the difference between the central pixel and the surrounding pixels in the neighborhood of the image. Besides,
is the potential energy of the pixel
,
is the energy value between the central pixel and other pixels in the neighborhood, and
are the eight neighborhoods of the pixel
. The energy function model is established as follows:
In Equation (20),
is the estimated value by the Gaussian curvature filter. In Equations (21)–(23),
is the edge-preserving potential function,
is the potential function between the central pixel and surrounding pixels,
is the kth smallest
in the neighborhood, and
is the sum of four smallest
values. Generally, the gray values of the noise-contaminated pixels are not continuous with that of the neighboring pixels, hence there will be a large variation in the gray values. In this paper, the specific assessment of the noise is performed using a 3 × 3 filter template to traverse the image, and then the energy value
of each pixel is calculated. When
is less than the threshold
, the pixel is considered as an uncontaminated signal point, otherwise the pixel is considered as a point contaminated by noise, and then the noise should be filtered. Therefore, the proposed energy function is used to distinguish the noise from the effective signal point of the star, and the superimposed noise both in the star point and the sky background are processed by the curvature filter, which can remove the noise without destroying the effective signal. As shown in
Figure 6a, the central pixel is a speckle noise, and the gray value changes from 30 to 13 after being processed, which indicates that the noise is denoised successfully. In addition, the central pixel in
Figure 6b is the contaminated star point, and the gray value changes from 46 to 28 after being processed, which achieves the goal of restoring the energy of the star.
4. Restoration of Motion Blurred Star Image
The restoration of a blurred star image requires the establishment of a degenerate/restoration model of the image. Gonzalez [
13] proposed that the degradation process of an image could be modeled using a degradation function
and an additive noise term
. An input image
is processed to produce a degraded image
. If
, the degradation function
and the additive noise term
are known, then an estimated
of the original image can be obtained. The degradation/restoration model of the image is shown in
Figure 7.
In this paper, RL algorithm is used to restore the blurred star image. The RL algorithm is a classical algorithm for image restoration, which assumes that the image obeys Poisson distribution and can be estimated by the maximum likelihood method. It is an iterative algorithm based on Bayesian analysis [
14], which requires little prior knowledge and provides good restoration effect. Its iterative equation is presented in Equation (24).
In Equation (24), represents the convolution operation, represents the correlation operation, and is a blurred star image. is the convolution operation result of Equations (12) and (15), where . and are the reconstructed images after k and k + 1 iterations, respectively.
It can be seen from Equation (24) that there is no termination condition in the iterative process of the RL algorithm, so the iteration number
must be selected based on experience. If
is too small, the blurred star image will not be fully restored and the result of the star location cannot meet the requirement of high accuracy. In contrast, if
is too large, the processing time of the blurred star image will increase, and it not only reduces the update rate of the star sensor, but also causes the amplification of noise in the iteration process, which seriously affects the accuracy of the star centroid [
15]. To solve the problem, we establish the termination condition of the iterative equation by using the star centroid coordinates in three consecutive frames of restored images and the parameter of the blurred star image.
In Equation (25), , , and are the star centroid coordinates in the restored images after , and iterations. and are the blurred length in the x and y directions, respectively. and are the functions about and , which should be selected appropriately after weighing the efficiency and effect of restoration. By using the proposed algorithm, the iterative process can be stopped once the restored star image satisfies the termination condition.