A fast restoration method for spaceborne remote sensing images

Abstract

The present invention belongs to the field of spaceborne remote sensing image processing technology and provides a method for rapid restoration of spaceborne remote sensing images, comprising: S1: randomly sampling an original remote sensing image g(x,y) to cut out a sample g1(x,y) of size M*M (pixels); S2: performing rapid 8-bit quantization; S3: setting a threshold for a blur scale d, comparing the blur scale d of the quantized image sample calculated using an autocorrelation equation with the threshold, and if the blur scale d exceeds the threshold, executing step S4; and S4: restoring the image using an improved L-R filtering algorithm. The present invention uses a random method to obtain a small-scale sample from the original image, performs 8-bit quantization in parallel, and rapidly calculates the blur scale to determine whether restoration is required, thereby reducing computational time; and uses an improved L-R filtering for image restoration, simultaneously considering both no-reference evaluation parameters and reference evaluation parameters, thereby achieving higher image restoration quality.

Description

Technical Field

[0001] The invention belongs to the technical field of satellite-borne remote sensing image processing, and in particular relates to a method for quickly restoring satellite-borne remote sensing images. Background Art

[0002] Image restoration and reconstruction are crucial processes in remote sensing image processing. Generally speaking, remote sensing image degradation is caused by various factors, such as atmospheric disturbances, camera vibration, target radiation, and electronic signal conversion. The goal of remote sensing image restoration is to restore degraded remote sensing images to their original state as closely as possible.

[0003] Traditional image restoration models generally use inverse filtering methods. Since most image information exists at the edges of the image, using filtering methods for image restoration must not only remove most blur and noise but also preserve image details.

[0004] Because image detail and noise often coexist in the same area, image denoising and smoothing, while preserving detail, become conflicting issues. Wiener filtering can resolve this conflict to a certain extent, but the calculation of the power spectrum ratio of noise to signal is quite complex. The Lucky-Richardson (LR) algorithm is a typical nonlinear method that can achieve good restoration results even when noise information is unknown. However, the point spread function of the LR filter must be known, making it unsuitable for spaceborne applications. Summary of the Invention

[0005] In order to solve the problems existing in the background technology and realize the rapid restoration of satellite remote sensing images, the present invention proposes a method for rapid restoration of satellite remote sensing images. The specific technical solution is as follows:

[0006] A method for rapidly restoring satellite-borne remote sensing images comprises the following steps:

[0007] S1: Randomly sample the original remote sensing image g(x,y) and cut out a sample g1(x,y) of size M*M (pixels);

[0008] S2: Perform fast 8-bit quantization;

[0009] S3: Setting a threshold value of the blur scale d, comparing the blur scale d of the quantized image sample calculated by the autocorrelation equation with the threshold value, and if it exceeds the threshold value, executing step S4;

[0010] S4: Restoration is performed using the improved LR filtering algorithm, including the following steps:

[0011] Set the minimum signal-to-noise ratio η, the allowed grayscale average gradient θ and the maximum number of iterations T as termination conditions;

[0012] The first time the signal-to-noise ratio is calculated, the original image is Gaussian filtered and used as the reference image for iterative calculation of the signal-to-noise ratio;

[0013] During the iteration process, if the signal-to-noise ratio is greater than the set threshold, Gaussian filtering is also performed.

[0014] Furthermore, the specific process of step S2 includes:

[0015] The 8-bit quantization process is calculated in parallel. If the quantization bit of the current remote sensing image is K, K ≥ 8, then the transformation equation of each pixel Pix(i, j) is:

[0016]

[0017] Pix s (i, j) represents the pixel value of the sample g1(x, y) at the pixel point (i, j);

[0018] The autocorrelation equation is constructed by the linear motion equation. The differential equation is quickly calculated in the horizontal and vertical directions for the quantized samples to obtain the autocorrelation equation of each row. The mean of all autocorrelation functions is used to estimate the motion blur scale d. The point spread function is:

[0019]

[0020] Furthermore, the specific process of step S4 includes:

[0021] S41: Initialize T, η and θ, where T is the maximum number of iterations, η is the minimum signal-to-noise ratio, and θ is the allowed grayscale average gradient. The initial iteration equation is:

[0022] r0(x,y)=g(x,y);

[0023] Initial iteration number t = 0;

[0024] S42: Perform Gaussian filtering on the original image g(x,y) to obtain g G (x,y);

[0025] S43: Execute LR filtering operation. The LR filtering equation is:

[0026]

[0027] S44: Calculate the grayscale mean gradient GMG of the filtered image t , signal-to-noise ratio SNR t , the current iteration number t = t + 1; the equation for calculating the signal-to-noise ratio is:

[0028]

[0029] Where w and h represent the width and height of the image respectively.

[0030] S45: If t=T or GMGt=θ, end the iteration process; otherwise, go to step S46;

[0031] S46: Comparing SNR t and η, if SNR t >η, return to S43; otherwise, for r t (x,y) is Gaussian filtered to get the new g G (x,y), return to S43.

[0032] The beneficial effects achieved by the present invention are:

[0033] First, the present invention uses a random method to obtain small-scale samples from the original image, uses a parallel method to perform 8-bit quantization, and quickly calculates the blur scale to determine whether restoration is required, thereby reducing the operation time;

[0034] Second, an improved LR filter is used for image restoration, which takes into account both no-reference evaluation parameters and reference evaluation parameters, and can achieve higher image restoration quality;

[0035] Third, it can be deployed on satellite-borne CPUs, smart chips, etc., with a short processing time, and can be used for real-time restoration of satellite-borne remote sensing images and image restoration of remote sensing quick-view systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of spaceborne remote sensing image restoration.

[0037] Figure 2 This is a comparison diagram of the effects before and after the restoration treatment in Example 1. DETAILED DESCRIPTION

[0038] To facilitate those skilled in the art to understand the present invention, specific implementations of the present invention are described below with reference to embodiments and drawings.

[0039] The present invention is a method for rapid restoration of satellite-borne remote sensing images. The main process is as follows: Figure 1As shown in the figure, the method first randomly samples the original remote sensing image g(x,y), cuts out a sample g1(x,y) of size M*M (pixels), and performs fast 8-bit quantization. The blur scale d of the quantized image sample is calculated using the autocorrelation equation, and the need for restoration is determined based on the set threshold. If restoration is required, it is restored using an improved LR filter, with the minimum signal-to-noise ratio η, the allowed grayscale average gradient θ, and the maximum number of iterations T as termination conditions. When calculating the signal-to-noise ratio for the first time, the original image is Gaussian filtered and used as the reference image for calculating the signal-to-noise ratio. If the signal-to-noise ratio during the iteration process is greater than the set threshold, Gaussian filtering is also performed.

[0040] Example 1,

[0041] This embodiment uses Figure 2 Taking the 4096*4096 image shown in the figure as an example, the restoration process of the present invention is introduced. The main implementation steps are as follows:

[0042] (1) After receiving the original remote sensing image g(x,y), the onboard processing unit randomly selects a 256*256 slice g1(x,y) from the original image and then performs 8-bit fast quantization. If the processing hardware is a CPU, multi-threaded concurrent processing is used. If the processing hardware is a GPU or intelligent processing chip, CUDA or OpenCL concurrent processing is used to improve processing speed.

[0043] (2) Calculate the fuzzy scale d of g1(x,y) through the autocorrelation equation to determine whether restoration operation is required.

[0044] The calculation method is as follows:

[0045] Step 1: Perform a differential operation on the degraded image slice g1(x,y) in the vertical direction.

[0046] g y (i,j)=g(i,j)-g(i,j-1)

[0047] Step 2: Perform differential operation on the degraded image slice g1(x,y) in the horizontal direction.

[0048] g yx (i,j)=g y (i,j)-g y (i-1,j)

[0049] Step 3: Calculate the autocorrelation function of each row of the degraded image. If there are l pixels in the jth row, the autocorrelation function can be expressed as:

[0050]

[0051] Where k is the row number and M is the width of the image.

[0052] Step 4: Calculate the mean of all autocorrelation functions and draw the autocorrelation function curve. Calculate the minimum value of the curve to estimate the degradation scale of the degraded image.

[0053] LR filtering is used to restore remote sensing images. The implementation steps are as follows:

[0054] Step 1: Initialize T, η, and θ, where T is the maximum number of iterations, η is the minimum signal-to-noise ratio, and θ is the allowed grayscale average gradient. The initial iteration equation is r0(x, y) = g(x, y). The initial number of iterations is t = 0.

[0055] Step 2: Perform Gaussian filtering on the original image g(x,y) to obtain g G (x,y);

[0056] Step 3: Perform LR filtering operation. The LR filtering equation is:

[0057]

[0058] Step 4: Calculate the grayscale mean gradient GMG of the filtered image t , signal-to-noise ratio SNR t , the current iteration number t=t+1. The equation for calculating the signal-to-noise ratio of the present invention is:

[0059]

[0060] Where w and h represent the width and height of the image respectively.

[0061] The calculation equation of the grayscale average gradient is as follows:

[0062]

[0063] Step 5: If t = T or GMG t =θ, the iteration process ends. Otherwise, go to step 6; Step 6: Compare SNR t and η, if SNR t >η, return to step 3; otherwise, t (x,y) is Gaussian filtered to get the new g G (x,y), return to step 3.

[0064] This embodiment uses the improved LR filter to restore the image, taking into account both the no-reference evaluation parameters and the reference evaluation parameters. The restoration results are shown in Figure 2 As shown, higher image restoration quality is achieved.

[0065] The above embodiments of the present invention do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for rapid restoration of spaceborne remote sensing images, characterized in that: It includes the following steps: S1: Randomly sample the original remote sensing image g(x,y) and cut out a sample g1(x,y) of M*M pixels; S2: Perform fast 8-bit quantization; S3: Setting a threshold value of the blur scale d, comparing the blur scale d of the quantized image sample calculated by the autocorrelation equation with the threshold value, and if it exceeds the threshold value, executing step S4; S4: Restoration is performed using the improved LR filtering algorithm, including the following steps: Set the minimum signal-to-noise ratio η, the allowed grayscale average gradient θ and the maximum number of iterations T as termination conditions; The first time the signal-to-noise ratio is calculated, the original image is Gaussian filtered and used as the reference image for iterative calculation of the signal-to-noise ratio; During the iteration process, if the signal-to-noise ratio is greater than the set threshold, Gaussian filtering is also performed; The specific process of step S2 includes: The 8-bit quantization process is calculated in parallel. If the quantization bit of the current remote sensing image is K, K ≥ 8, then the transformation equation of each pixel Pix(i, j) is: Pix s (i, j) represents the pixel value of the sample g1(x, y) at the pixel point (i, j); The autocorrelation equation is constructed by the linear motion equation. The differential equation is quickly calculated in the horizontal and vertical directions for the quantized samples to obtain the autocorrelation equation of each row. The mean of all autocorrelation functions is used to estimate the motion blur scale d. The point spread function is: The specific process of step S4 includes: S41: Initialize T, η and θ, where T is the maximum number of iterations, η is the minimum signal-to-noise ratio, and θ is the allowed grayscale average gradient. The initial iteration equation is: r0(x,y)=g(x,y); Initial iteration number t=0; S42: Perform Gaussian filtering on the original image g(x,y) to obtain g G (x,y); S43: Execute LR filtering operation. The LR filtering equation is: S44: Calculate the grayscale mean gradient GMG of the filtered image t , signal-to-noise ratio SNR t , the current iteration number t = t + 1; The equation for calculating the signal-to-noise ratio is: Among them, w and h represent the width and height of the image respectively; S45: If t=T or GMGt=θ, end the iteration process; otherwise, go to step S46; S46: Comparing SNR t and η, if SNR t >η, return to S43; otherwise, for r t (x,y) is Gaussian filtered to get the new g G (x,y), return to S43.

Images