Method and System for Centroid Location of Space Debris Based on Lightweight Super-Resolution

Abstract

The present invention discloses a method and system for centroid positioning of space debris based on lightweight super-resolution, which relates to the technical field of centroid positioning. The specific steps are as follows: obtaining data to be processed; inputting the data to be processed into a deep convolutional network for calculation, respectively inputting the calculated data into a first channel and a second channel for three-level upsampling processing, and performing mean summation on the upsampled data output by the first channel and the second channel to obtain mean data; inputting the mean data into the deep convolutional network to obtain super-resolution data; performing centroid solution on the super-resolution data to obtain positioning data. The present invention designs a lightweight super-resolution deep convolutional network Tpf-Net. By establishing an upsampling and downsampling dual-channel parallel fusion super-resolution network and combining a five-layer pyramid convolutional network, 8-fold super-resolution data is obtained after three upsamplings, and high-precision positioning calculation can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of centroid positioning, and more specifically, to a method and system for centroid positioning of space debris based on lightweight super-resolution. Background Art

[0002] With the continuous growth of global space activities, the problem of space debris has become a key factor affecting the safety of spacecraft and the sustainability of space exploration. Space debris is mainly composed of failed satellites, rocket remnants, and debris generated by collisions of tiny objects. They orbit the Earth at high speeds, posing a serious threat to normal operating satellites and other space assets. Once a collision occurs, it may lead to the loss of expensive space assets and even threaten the lives of astronauts. Therefore, effectively monitoring and managing space debris has become an important task for ensuring the long-term safety of the space environment.

[0003] Traditional imaging technologies are limited by resolution and are difficult to accurately identify and locate these space debris of different sizes and shapes. Super-resolution imaging technology, through advanced image processing algorithms, can break through the resolution limit of the physical optical system and provide clearer and more detailed images than the original acquired data. This ability is crucial for accurately tracking and analyzing space debris.

[0004] However, imaging of space debris faces a series of unique challenges:

[0005] Long distance: The distance between space debris and the observation equipment is usually very far, resulting in a relatively small effective area of the target reflection.

[0006] Low reflectivity: Many debris surfaces may not have good reflection characteristics, making their contrast in the image very low.

[0007] Large background interference: The space background is complex and variable, including stars, planets, and other artificial or natural celestial bodies, which may all appear as interference sources.

[0008] Few effective pixels: Especially for small debris, the number of effective pixels that can be obtained under conventional imaging conditions is extremely limited, which directly affects the centroid positioning accuracy.

[0009] Low signal-to-noise ratio: Due to weak signal strength and a large amount of noise, it is very difficult to achieve high-precision centroid positioning.

[0010] Therefore, how to improve the super-resolution processing efficiency and imaging quality during the centroid positioning of space debris is an urgent problem for those skilled in the art. Summary of the Invention

[0011] In view of this, the present invention provides a method and system for centroid localization of space debris based on lightweight super-resolution, which overcomes the above-mentioned defects.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] A method for centroid localization of space debris based on lightweight super-resolution, the specific steps are as follows:

[0014] Obtain the data to be processed;

[0015] Input the data to be processed into a deep convolutional network for calculation, input the calculated data into the first channel and the second channel for three-level upsampling processing respectively, and perform mean summation on the upsampled data output by the first channel and the second channel to obtain mean data;

[0016] Input the mean data into the deep convolutional network to obtain super-resolution data;

[0017] Solve the centroid of the super-resolution data to obtain the localization data.

[0018] Optionally, the specific steps for performing three-level upsampling processing in the first channel are: perform three upsamplings on the calculated data in sequence to obtain the first upsampled data, the second upsampled data, and the third upsampled data respectively.

[0019] Optionally, the specific steps for performing three-level upsampling processing in the second channel are: the first-level upsampling is: perform downsampling processing on the calculated data to obtain downsampled data; perform upsampling processing on the downsampled data to obtain the fourth upsampled data, and perform mean summation on the fourth upsampled data and the first upsampled data in the same dimension to obtain the first mean data; the second-level upsampling is: perform upsampling processing on the first mean data to obtain the fifth upsampled data, and perform mean summation on the fifth upsampled data and the second upsampled data to obtain the second mean data; the third-level upsampling is: perform upsampling processing on the second mean data to obtain the sixth upsampled data, and perform mean summation on the sixth upsampled data and the third upsampled data to obtain the third mean data.

[0020] Optionally, before performing upsampling or downsampling processing, the data needs to be calculated through the deep convolutional network.

[0021] Optionally, the deep convolutional network adopts a pyramid convolutional network.

[0022] A system for centroid localization of space debris based on lightweight super-resolution, comprising:

[0023] A data acquisition module for obtaining the data to be processed;

[0024] A multi-level upsampling module for inputting the data to be processed into a deep convolutional network for calculation, inputting the calculated data into a first channel and a second channel respectively for three-level upsampling processing, and performing mean summation on the upsampled data output from the first channel and the second channel to obtain mean data;

[0025] A super-resolution data acquisition module for inputting the mean data into the deep convolutional network to obtain super-resolution data;

[0026] A position acquisition module for solving the centroid of the super-resolution data to obtain positioning data.

[0027] Optionally, the multi-level upsampling module includes a first calculation unit, a first channel processing unit, and a second channel processing unit;

[0028] The first calculation unit is used for calculating the data to be processed by using the deep convolutional network to obtain the calculated data;

[0029] The first channel processing unit is used for performing three-level upsampling processing on the calculated data to obtain first-channel upsampled data; the first-channel upsampled data includes: first upsampled data, second upsampled data, and third upsampled data;

[0030] The second channel processing unit is used for performing downsampling on the calculated data and then performing three-level upsampling processing to obtain second-channel upsampled data, and performing mean summation on the first-channel upsampled data and the second-channel upsampled data to obtain the mean data.

[0031] Optionally, the second channel processing unit includes:

[0032] A downsampling sub-unit for performing downsampling processing on the calculated data to obtain downsampled data;

[0033] A first upsampling sub-unit for performing upsampling processing on the downsampled data to obtain fourth upsampled data, and the fourth upsampled data is in the same dimension as the first upsampled data;

[0034] A first mean summation sub-unit for performing mean summation on the fourth upsampled data and the first upsampled data to obtain first mean data;

[0035] A second upsampling sub-unit for performing upsampling processing on the first mean data to obtain fifth upsampled data;

[0036] A second mean summation sub-unit for performing mean summation on the fifth upsampled data and the second upsampled data to obtain second mean data;

[0037] A third upsampling sub-unit, configured to perform upsampling processing on the second mean data to obtain sixth upsampled data;

[0038] A third mean summation sub-unit, configured to perform mean summation on the sixth upsampled data and the third upsampled data to obtain third mean data;

[0039] A calculation sub-unit, configured to perform calculation on the data before upsampling or downsampling processing by using the deep convolutional network.

[0040] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a method and system for centroid positioning of space debris based on lightweight super-resolution, designs a lightweight super-resolution deep convolutional network Tpf-Net, and through establishing an upsampling and downsampling dual-channel parallel fusion super-resolution network, combining a five-layer pyramid convolutional network, obtains 8-fold super-resolution data through three times of upsampling, and can achieve high-precision positioning calculation. In the calculation process, only three calculations including upsampling, downsampling, pyramid convolution and mean summation are included, and the overall architecture is simple and suitable for parallelism. Each calculation process can adopt multiple calculation operators in parallel, or multiple convolutions in parallel, greatly improving the data processing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0042] Figure 1 It is a schematic flow chart of the method provided by the present invention;

[0043] Figure 2 It is a schematic diagram of the architecture of the lightweight super-resolution deep convolutional network provided by the present invention;

[0044] Figure 3 It is a schematic diagram of the pyramid convolution structure provided by the present invention;

[0045] FIG. 4(a) is a schematic diagram of low signal-to-noise ratio non-tail data in the test data set provided by the present invention; FIG. 4(b) is a schematic diagram of high signal-to-noise ratio non-tail data in the test data set provided by the present invention; FIG. 4(c) is a schematic diagram of low signal-to-noise ratio tail data in the test data set provided by the present invention; FIG. 4(d) is a schematic diagram of high signal-to-noise ratio tail data in the test data set provided by the present invention;

[0046] Figure 5Schematic diagram of centroid positioning accuracy of different algorithms under different signal-to-noise ratios provided by the present invention. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] The present invention discloses a method and system for centroid positioning of space debris based on lightweight super-resolution. By constructing a super-resolution network that includes upsampling and downsampling dual-channel parallel fusion and combining it with a five-layer pyramid convolution network, 8-fold super-resolution data can be generated through three upsamplings. On this basis, the centroid of the super-resolution data is solved to achieve the goal of high-precision positioning calculation, as Figure 1 shown. The specific steps are as follows:

[0049] Step 1: Obtain the data to be processed;

[0050] Step 2: Input the data to be processed into the deep convolutional network for calculation, input the calculated data into the first channel and the second channel respectively for three-level upsampling processing, and perform mean summation on the upsampled data output by the first channel and the second channel to obtain mean data;

[0051] Step 3: Input the mean data into the deep convolutional network to obtain super-resolution data;

[0052] Step 4: Solve the centroid of the super-resolution data to obtain positioning data.

[0053] Furthermore, the above methods are all implemented based on a lightweight super-resolution deep convolutional network (Tpf-Net), and its structure is as Figure 2 shown, including: a deep convolutional network layer (conv), an upsampling layer, and an accumulation summation layer. The method implemented by the entire Tpf-Net is roughly divided into 3 steps: input an image not greater than , enter the deep convolutional network (conv) for calculation, and the results are simultaneously upsampled and downsampled to obtain the first upsampled data of and the downsampled data of ; the first upsampled data and the downsampled data of enter the deep convolutional network (conv) for calculation respectively, and the data of the same dimension output by the two channels are subjected to mean summation calculation to obtain ; Then perform upsampling, and finally when the upsampling reaches , the result of mean summation passes through a deep convolutional network (conv) to obtain the final super-resolution data .

[0054] Further, the calculation process of the first upsampled data and the downsampled data is as follows:

[0055] (1);

[0056] Among them, and are the upsampling and downsampling functions respectively, is to perform on the image pyramid convolution calculation. In this embodiment, is 32.

[0057] In one embodiment, the specific steps for performing three-level upsampling processing in the first channel are: successively perform upsampling on the calculated data three times to obtain the first upsampled data, the second upsampled data, and the third upsampled data respectively.

[0058] In one embodiment, the specific steps for performing three-level upsampling processing in the second channel are: the first-level upsampling is: perform downsampling on the calculated data to obtain the downsampled data; perform upsampling on the downsampled data to obtain the fourth upsampled data, and perform mean summation of the fourth upsampled data and the first upsampled data in the same dimension to obtain the first mean data; the second-level upsampling is: perform upsampling on the first mean data to obtain the fifth upsampled data, and perform mean summation of the fifth upsampled data and the second upsampled data to obtain the second mean data; the third-level upsampling is: perform upsampling on the second mean data to obtain the sixth upsampled data, and perform mean summation of the sixth upsampled data and the third upsampled data to obtain the third mean data.

[0059] In one embodiment, before performing upsampling or downsampling processing, the data needs to be calculated through a deep convolutional network, and the calculation process is:

[0060] (2);

[0061] In the formula, is the first upsampled data; is the second upsampled data; is the pyramid convolution network; is the first upsampled data of the second channel; is the downsampled data; is the third upsampling data; is the fourth upsampling data; is the fifth upsampling data; is the first mean data; is the sixth upsampling data; is the second mean data;

[0062] The mean summation process of the upsampling results at each level of the dual-channel is as follows:

[0063] (3);

[0064] In the formula, is the third mean data; represents taking the average value.

[0065] The super-resolution data is:

[0066] (4);

[0067] In one embodiment, the deep convolutional network adopts a pyramid convolutional network.

[0068] Furthermore, the deep convolutional network (conv) is a five-layer pyramid convolutional network (Pyramid convolution, ), and its architecture is as Figure 3 shown. The same convolutional calculation structure is adopted for different input pixel sizes. Specifically, for an image with an input of , first perform 16 convolution kernel calculations, where , to obtain 16 subgraphs; then pass through 32 groups of 16 convolutions to obtain 32 subgraphs; then pass through 16 groups of 32 convolutions to obtain 16 subgraphs; finally, pass through 16 groups of convolutions to obtain the convolved image. The pyramid and inverted pyramid cascaded convolutional network is used to calculate the upsampled and downsampled images, enabling accurate restoration of the upsampled and downsampled images.

[0069] Based on the above description, it is not difficult to see that Tpf-Net is particularly suitable for digital logic acceleration calculations, mainly as follows:

[0070] I. Small computational complexity: The pyramid convolution process is only a five-layer convolution, and the computational complexity reaches the highest when convolving an image of , only requiring less than weights ( Convolution), and the total network weights do not exceed , which is 90% lower than that of traditional super-resolution networks;

[0071] II. Fixed structure: The entire computational process only includes three operations: upsampling, pyramid convolution, and mean summation. The overall architecture is simple and suitable for parallel processing. Multiple computing operators can be used in parallel for each computing process, or multiple convolutions can be used in parallel to improve real-time performance;

[0072] III. High parallelism: The eight pyramid convolution calculations are independent of each other. The intermediate variables are only the input and output of the previous and subsequent time series, and do not affect the sequence of convolution calculations within each pyramid convolution, which is conducive to improving the parallelism.

[0073] In an embodiment, the above method is verified. The specific steps are as follows:

[0074] Construct a dataset: Use the actual space debris monitoring public data of the ground-based wide-angle camera array (GWAC) to create a dataset for super-resolution calculation, including 800 training images and 200 test image pairs. The Tpf-Net model is used to train the data and test the test dataset. At the same time, for test comparison, the improved least squares estimation algorithm, point spread model algorithm, and iterative adaptive window algorithm are selected as the comparison algorithms in this embodiment to test the same test dataset. As three commonly used mainstream centroid localization algorithms, their centroid localization accuracy performance has a certain representativeness.

[0075] Model setting: In this embodiment, three types of Tpf-Net are designed and tested according to different convolution kernel sizes. Among them, The convolution kernel version is the basic version of Tpf-Net, The convolution kernel version is Tpf-Net+, The convolution kernel version is Tpf-Net++. Three versions of convolution kernel sizes are used to test the impact of convolution kernel size on the network effect.

[0076] Training configuration: Process a dataset of 1000 original images, and use the sampling filtering method to obtain 8-fold low-resolution data. 800 pairs of training images are trained using the Pytorch framework. The training platform uses NVIDIA's GPU3060Ti, and the processor uses I7-10700.

[0077] Comparison of experimental results of the dataset: As shown in Table 1, it is compared with the previous three types of centroid localization methods for space debris and stellar targets, and the accuracy and computational complexity are also compared with the previous three commonly used super-resolution algorithms, including traditional computational models and those based on CNN. Traditional algorithms are mainly divided into two types: based on the point spread function and interpolation filtering. Among them, EIWA reaches the highest centroid localization accuracy of 0.0056 pixels. All super-resolution results are tested and compared by PSNR and SSIM.

[0078] Table 1 gives the quantitative comparison obtained by Tpf-Net with other super-resolution algorithms and centroid localization algorithms, as well as the comparison of the finally obtained centroid localization accuracy.

[0079] Table 1 Comparison of centroid localization methods and super-resolution algorithms under different signal-to-noise ratios

[0080]

[0081] The first three algorithms use traditional computational methods for high-precision centroid localization, so there are no values of PNSR and SSIM. The least squares method is used to predict the centroid position, and the error is large when the signal-to-noise ratio is low, reaching below 0.2 pixels. The point spread function method can reach about 0.05 pixels at low signal-to-noise ratios, and the improvement of the localization accuracy by increasing the signal-to-noise ratio is not high. The adaptive window method based on energy iteration reaches 0.02 pixels at low signal-to-noise ratios and about 0.005 at high signal-to-noise ratios.

[0082] For the middle three methods, the super-resolution function is realized by using the iterative and function reconstruction network, the enhanced octave convolution network, and the multi-model integration network respectively. Among them, EOctConv and MMSR are difficult to form training convergence at low signal-to-noise ratios, which is mainly related to the fact that their networks include multiple types of architectures. It can be seen that the previous algorithms based on CNN are difficult to achieve high centroid localization accuracy because the super-resolution intelligent methods are mainly applicable to structured images or targets, and it is difficult to obtain good results for the point target information of space debris.

[0083] The three network versions designed in this embodiment have similar centroid localization accuracies obtained by training at low signal-to-noise ratios, and the convolution kernel version has higher accuracy at high signal-to-noise ratios. However, considering that the weight of the convolution kernel increases by 4 times, and the centroid localization accuracy increases by less than 20%. Moreover, most targets in the field of space debris are in a low signal-to-noise ratio state, so the Tpf-Net version has higher efficiency.

[0084] Comparison of experimental results of different types of data: For 200 test image data, they can be divided into four categories, 50 for each category: including low signal-to-noise ratio without trailing, high signal-to-noise ratio without trailing, low signal-to-noise ratio with trailing, and high signal-to-noise ratio with trailing, as shown in Figures 4(a)-4(d). Due to the differences in the distribution characteristics of these four types of data, and whether there is trailing will have a great impact on the centroid positioning accuracy, so the results of testing with these four types of data are compared and analyzed, as shown in Table 2. Among them, trailing less than 4 pixels is considered without trailing, and the trailing length of the trailing data is distributed between 5 and 10 pixels. The case of trailing exceeding 10 pixels is not considered here.

[0085] Among them, the case of low signal-to-noise ratio without trailing is the same as the case where the signal-to-noise ratio < 6, and the test results are already available in Table 1. The accuracy of the test results of the trailing data in Table 2 is reduced by one order of magnitude compared with the results without trailing, and the centroid positioning accuracy of the low signal-to-noise ratio trailing data is even reduced to 0.1 pixel. This also shows that for the low signal-to-noise ratio trailing data, due to the "break" of the target as a whole caused by the trailing, it is difficult to identify the target body well and perform super-resolution prediction during the super-resolution calculation process, resulting in low centroid positioning accuracy. Using morphological preprocessing can effectively improve the centroid positioning accuracy of the "break" of the low signal-to-noise ratio trailing. Through the pyramid convolution of upsampling and downsampling, good results can be obtained for a single image. The demarcation threshold between high signal-to-noise ratio and low signal-to-noise ratio is 6.

[0086] Table 2 Comparison of centroid positioning methods for different types of test data and super-resolution algorithm based on progressive preprocessing

[0087]

[0088] In the case of trailing, Tpf-Net++ has a higher centroid positioning accuracy than Tpf-Net. This is because when the trailing length is between 5 and 10, the increase in the convolution kernel size can improve the perception ability for the trailing data, thus generating a better super-resolution image and improving the positioning accuracy.

[0089] Comparison of experimental results of data with different signal-to-noise ratios: The data with different signal-to-noise ratios in the dataset are tested, and the comparison of the centroid positioning accuracy results under different algorithms is obtained, as Figure 5 shown.

[0090] Generally speaking, the increase in positioning accuracy of various algorithms tends to stabilize when the signal-to-noise ratio (SNR) is above 4. On the one hand, it is because faint targets that have a greater impact on accuracy have better algorithm adaptability at higher SNRs and can synchronously improve positioning accuracy with the increase in SNR. On the other hand, it is because for most algorithms, the data with an SNR below 3 is overly affected by noise, making it difficult for target information to contribute to accuracy. Among them, the least squares method has relatively poor overall accuracy because it is only a simple least squares estimation of the target area and its surroundings. The point spread function method is difficult to achieve high-precision indicators at low SNRs. The EIWA algorithm uses the method of window function iteration to obtain relatively high accuracy. Compared with the lightweight super-resolution network Tpf-Net designed in this paper, the accuracy difference is not significant at high SNRs. At low SNRs, Tpf-Net improves the accuracy by 40% compared with EIWA.

[0091] In this embodiment, a lightweight super-resolution deep convolutional network Tpf-Net is designed. By establishing an upsampling and downsampling dual-channel parallel fusion super-resolution network and combining it with a five-layer pyramid convolutional network, 8-fold super-resolution data is obtained after three upsamplings. The centroid of the obtained super-resolution data is solved to achieve high-precision positioning calculation. Compared with previous algorithms, the key points in the calculation process are the upsampling and downsampling calculations and the pyramid convolutional calculations before and after sampling, and the weight is only less than , with a simple architecture, suitable for digital logic acceleration, and suitable for on-orbit lightweight computing applications. During the experiment, comprehensive tests were carried out on data of different types and SNRs in GWAC. Compared with previous algorithms, the centroid positioning accuracy is improved by more than 40% at low SNRs, and the positioning accuracy is improved by 10 times at high SNRs. By comparing the accuracy with previous algorithms, the effectiveness of Tpf-Net is demonstrated.

[0092] On the other hand, this embodiment also discloses a space debris centroid positioning system based on lightweight super-resolution, which applies the method disclosed above and includes:

[0093] A data acquisition module for acquiring data to be processed;

[0094] A multi-level upsampling module for inputting the data to be processed into a deep convolutional network for calculation, inputting the calculated data into the first channel and the second channel for three-level upsampling processing, and performing mean summation on the upsampled data output from the first channel and the second channel to obtain mean data;

[0095] A super-resolution result acquisition module for inputting the mean data into a deep convolutional network to obtain super-resolution data;

[0096] A position acquisition module for solving the centroid of the super-resolution data to obtain positioning data.

[0097] In one embodiment, the multi-level upsampling module includes a first computing unit, a first channel processing unit, and a second channel processing unit;

[0098] The first computing unit is configured to perform calculations on the data to be processed using a depth convolutional network to obtain calculation data;

[0099] The first channel processing unit is configured to perform three-level upsampling on the calculation data to obtain first-channel upsampled data; the first-channel upsampled data includes: first upsampled data, second upsampled data, and third upsampled data;

[0100] The second channel processing unit is configured to perform downsampling on the calculation data and then perform three-level upsampling to obtain second-channel upsampled data, and perform mean summation on the first-channel upsampled data and the second-channel upsampled data to obtain mean data.

[0101] In one embodiment, the second channel processing unit includes:

[0102] The downsampling subunit is configured to perform downsampling on the calculation data to obtain downsampled data;

[0103] The first upsampling subunit is configured to perform upsampling on the downsampled data to obtain fourth upsampled data, and the fourth upsampled data is in the same dimension as the first upsampled data;

[0104] The first mean summation subunit is configured to perform mean summation on the fourth upsampled data and the first upsampled data to obtain first mean data;

[0105] The second upsampling subunit is configured to perform upsampling on the first mean data to obtain fifth upsampled data;

[0106] The second mean summation subunit is configured to perform mean summation on the fifth upsampled data and the second upsampled data to obtain second mean data;

[0107] The third upsampling subunit is configured to perform upsampling on the second mean data to obtain sixth upsampled data;

[0108] The third mean summation subunit is configured to perform mean summation on the sixth upsampled data and the third upsampled data to obtain third mean data;

[0109] The second computing subunit is configured to perform calculations on the data before upsampling or downsampling using a depth convolutional network.

[0110] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the apparatuses disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method part.

[0111] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A space debris centroid positioning method based on lightweight super-resolution, characterized in that: The specific steps are: Get the data to be processed; Input the data to be processed into a deep convolutional network for calculation, input the calculated data into the first channel and the second channel respectively for three-level upsampling processing, and average the upsampled data output by the first channel and the second channel to obtain the average data; wherein the first-level upsampling processing step of the second channel includes first downsampling the calculated data to generate downsampled data, and then upsampling the downsampled data; before downsampling, first-level upsampling, second-level upsampling and third-level upsampling, the data is calculated by the deep convolutional network; Inputting the mean data into the deep convolutional network to obtain super-resolution data; the deep convolutional network is a five-layer pyramid convolutional network, including a plurality of convolutional networks using pyramids and inverted pyramids in cascade, and performing convolution calculation on the input data; The centroid of the super-resolution data is solved to obtain positioning data.

2. The method for locating the centroid of space debris based on lightweight super-resolution according to claim 1, characterized in that: The specific steps of performing three-level upsampling processing in the first channel are: upsampling the calculated data three times in sequence to obtain first upsampling data, second upsampling data and third upsampling data respectively.

3. The method for locating the centroid of space debris based on lightweight super-resolution according to claim 2, characterized in that: The specific steps of performing three-level upsampling processing in the second channel are as follows: the first level upsampling is: performing downsampling processing on the calculated data to obtain the downsampling data; performing upsampling processing on the downsampling data to obtain fourth upsampling data, and performing mean summing of the fourth upsampling data and the first upsampling data in the same dimension to obtain first mean data; the second level upsampling is: performing upsampling processing on the first mean data to obtain fifth upsampling data, and performing mean summing of the fifth upsampling data and the second upsampling data to obtain second mean data; the third level upsampling is: performing upsampling processing on the second mean data to obtain sixth upsampling data.

4. A space debris centroid positioning system based on lightweight super-resolution, characterized in that: include: A data acquisition module, used to obtain data to be processed; A multi-level upsampling module, used for inputting the data to be processed into a deep convolutional network for calculation, inputting the calculated data into a first channel and a second channel respectively for three-level upsampling processing, and performing mean summation on the upsampled data output by the first channel and the second channel to obtain mean data; wherein the first-level upsampling processing step of the second channel includes first performing downsampling processing on the calculated data to generate downsampled data, and then performing upsampling processing on the downsampled data; before downsampling, first-level upsampling, second-level upsampling and third-level upsampling, the data is calculated by the deep convolutional network; A super-resolution data acquisition module, used for inputting the mean data into the deep convolutional network to obtain super-resolution data; the deep convolutional network is a five-layer pyramid convolutional network, including a plurality of convolutional networks using pyramids and inverted pyramids cascaded, and performing convolution calculation on the input data; The position acquisition module is used to solve the centroid of the super-resolution data to obtain positioning data.

5. The space debris centroid positioning system based on lightweight super-resolution according to claim 4, characterized in that: The multi-stage upsampling module includes a first calculation unit, a first channel processing unit, and a second channel processing unit; The first computing unit is used to compute the data to be processed using the deep convolutional network to obtain the computed data; The first channel processing unit is used to perform three-level upsampling processing on the calculated data to obtain first channel upsampling data; The first channel up-sampled data includes: first up-sampled data, second up-sampled data and third up-sampled data; The second channel processing unit is used to perform three-level upsampling processing on the calculated data to obtain second channel upsampling data, and to sum the first channel upsampling data with the second channel upsampling data to obtain the mean data.

6. The space debris centroid positioning system based on lightweight super-resolution according to claim 5, characterized in that: The second channel processing unit comprises: A down-sampling subunit, used for performing down-sampling processing on the calculated data to obtain the down-sampled data; A first up-sampling sub-unit is used to perform up-sampling processing on the down-sampled data to obtain fourth up-sampled data, where the fourth up-sampled data is in the same dimension as the first up-sampled data; a first mean summing subunit, configured to perform mean summing of the fourth up-sampled data and the first up-sampled data to obtain first mean data; A second up-sampling subunit is used to perform up-sampling processing on the first mean data to obtain fifth up-sampled data; a second mean summing subunit, configured to sum the fifth up-sampled data and the second up-sampled data to obtain second mean data; A third up-sampling subunit is used to perform up-sampling processing on the second mean data to obtain sixth up-sampled data; a third mean summing subunit, configured to perform mean summing of the sixth up-sampled data and the third up-sampled data to obtain third mean data; The computing subunit is used to perform computing on the data before upsampling or downsampling using the deep convolutional network.

Images