Satellite-borne massive image data distributed cooperative rapid high-precision processing system and method

Abstract

The application provides a kind of satellite-borne mass image data distributed cooperative fast high-precision processing system, comprising: imaging mode judging module, area matching calculation module, fast screening module and high-precision detection identification module;Imaging mode judging module judges the imaging mode of current work of remote sensing satellite, when remote sensing satellite works in the wide-area search imaging mode of ocean target, the original image data is input into fast screening module, when remote sensing satellite works in the point target detailed investigation imaging mode, the original image data is input into area matching calculation module;Fast screening module filters out the image suspected to contain suspicious target from original image data and inputs into high-precision detection identification module;Area matching calculation module constructs hot area, and the image belonging to hot area is filtered out from original image data and input into high-precision detection identification module;High-precision detection identification module carries out suspicious target detection and identification to the input image and outputs detection result.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of space remote sensing, and particularly relates to a kind of spaceborne mass image data distributed collaborative fast high-precision processing system and method. BACKGROUND

[0002] In recent years, a new generation of artificial intelligence technology represented by deep learning has been widely used in spaceborne data processing systems, and this kind of method has made great progress in performance compared with traditional methods. However, with the continuous improvement of spatial, temporal and spectral resolution of remote sensing satellites, the amount of data acquired by satellites is getting larger and larger (data rate has reached dozens of Gbps), and the processing rate of mainstream deep learning models (FPGA or AI chips) can only reach tens of Mbps (tens of frames) to hundreds of Mbps (hundreds of frames), which cannot meet the real-time high-precision processing requirements of spaceborne mass data.

[0003] Through big data analysis, it can be known that only a small part of data (less than 1%) contains real ship targets in the large amount of marine remote sensing image data acquired by remote sensing satellites, and most of them are pure ocean or cloud layer without target useless data. The current spaceborne processing system has the following shortcomings when detecting and processing ocean targets:

[0004] The entire processing flow needs to traverse all the images acquired by the satellite, that is, the specific processing flow divides the acquired original image into blocks, and no matter whether there is a target in the image block, each image block will be sent to the high-precision target detection and recognition network for processing; similarly, in the detection and recognition processing of fixed area targets such as airport aircraft and port ship, traversal processing is also performed; the existing method greatly wastes spaceborne computing resources and increases processing delay. SUMMARY

[0005] The technical problem solved by the present application is to overcome the shortcomings of the prior art and provide a kind of spaceborne mass image data distributed collaborative fast high-precision processing system and method, which effectively improves the efficiency and accuracy of satellite mass image data processing.

[0006] The technical solution of the present application is:

[0007] A kind of spaceborne mass image data distributed collaborative fast high-precision processing system, comprising: imaging mode judging module, area matching calculation module, fast screening module and high-precision detection and identification module, wherein:

[0008] An imaging mode judging module: judging the imaging mode of the remote sensing satellite currently working, when the remote sensing satellite works in the wide-area search imaging mode of the ocean target, inputting the original image data into the fast screening module, and when the remote sensing satellite works in the point target detailed imaging mode, inputting the original image data into the region matching calculation module;

[0009] A fast screening module: screening the image suspected to contain a suspicious target from the original image data and inputting the image into the high-precision detection and recognition module, and discarding the remaining images;

[0010] A region matching calculation module: constructing a hot region, screening the image belonging to the hot region from the original image data and inputting the image into the high-precision detection and recognition module, and discarding the remaining images;

[0011] A high-precision detection and recognition module: performing target detection and recognition on the input image and outputting a detection result.

[0012] Preferably, the fast screening module comprises one or more data processing units, each data processing unit adopts a fast screening network to screen the input image, the fast screening network comprises a first convolutional layer, a second convolutional layer, at least one feature extraction unit, a third convolutional layer, a fully connected layer and a Softmax layer arranged in sequence, the first convolutional layer comprises 32 1*1 convolutional kernels, the second convolutional layer comprises 64 3*3 convolutional kernels, the feature extraction unit comprises a multi-channel feature extraction unit, a Max pooling layer and a fourth convolutional layer arranged in sequence, the multi-channel feature extraction unit comprises a first channel, a second channel and a third channel, the first channel is connected with 64 3*3 convolutional kernels after 32 1*1 convolutional kernels, the second channel is connected with 64 1*1 convolutional kernels after 32 3*3 convolutional kernels, the third channel is 64 1*1 convolutional kernels, the fourth convolutional layer is 64 3*3 convolutional kernels, and the third convolutional layer comprises 32 3*3 convolutional kernels; the features output after convolution of the convolutional kernels in the first convolutional layer, the second convolutional layer, the first channel, the second channel, the third channel, the third convolutional layer or the fourth convolutional layer are activated by using a Relu activation function.

[0013] Preferably, the 1*1 convolutional kernel and the 3*3 convolutional kernel have a convolution step of 1.

[0014] Preferably, the number of the feature extraction units is 3-5, and the feature extraction units are connected in sequence.

[0015] Preferably, the high-precision detection and recognition module comprises a multi-modal data fusion unit, a high-resolution information maintaining unit, a main network and an auxiliary super-resolution learning branch, wherein:

[0016] The multi-modal data fusion unit: the original images are unified to the same resolution, and a fused image is formed by multi-source data fusion and input to the high-resolution information maintaining unit; the original images include a panchromatic image, a multi-spectral image and a near-infrared image;

[0017] The high-resolution information maintaining unit: details of the fused image are maintained and enhanced, and a high-resolution image with enhanced details is obtained and input to the backbone network;

[0018] The backbone network: low-level and high-level features of the high-resolution image with enhanced details are extracted, a high-precision detection and recognition model is used to judge the suspicious target category in the image, and the position and confidence of the suspicious target are calculated;

[0019] The auxiliary super-resolution learning branch: the low-level and high-level features of the image obtained by the backbone network are used for super-resolution learning, the local texture, detail features and high-level semantic information of the complex background image are obtained, and the high-precision detection and recognition model of the backbone network is corrected.

[0020] Preferably, the auxiliary super-resolution learning branch includes an encoder and a decoder, the encoder matches the low-level features of the image to the spatial size of the high-level features, and then combines the low-level features with the high-level features; the decoder decodes the information after the low-level features and the high-level features are combined, and outputs super-resolution features.

[0021] Preferably, the region matching calculation module constructs a hot region, including:

[0022] The earth's surface is divided into basic grids of equal size according to the latitude and longitude index, each basic grid is labeled and numbered, and one or more hot regions are formed according to the information on the ground, each hot region is composed of one or more basic grids.

[0023] Preferably, the region matching calculation module extracts images belonging to the hot region from the original image, including:

[0024] The original image is divided into image blocks of equal size, the ground latitude and longitude coordinates corresponding to the center point of the image block are calculated using the GPS and attitude information in the image auxiliary data, the basic grid number where the center point of the image block is located is determined according to the latitude and longitude coordinates, and then it is judged whether the basic grid belongs to the hot region, if it belongs, the image block is extracted, and if it does not belong, the image block is discarded.

[0025] Preferably, the fast screening module and the region screening module include a plurality of parallel data processing units.

[0026] A method for image processing by using a satellite-borne massive image data distributed collaborative fast high-precision processing system, including the following steps:

[0027] (1) receiving raw image data input by a satellite-borne camera;

[0028] (2) judging an imaging mode of satellite operation, when the satellite operates in a wide-area search imaging mode for ocean targets, entering step (3), and when the satellite operates in a point target detailed search imaging mode, entering step (5);

[0029] (3) dividing the raw image data into image blocks of equal size and inputting the image blocks into a fast screening module in sequence for suspected target screening processing, and entering step (4);

[0030] (4) when a suspected target is detected in an image block, inputting the image block into a high-precision detection and recognition module for recognition processing and outputting a detection result, otherwise discarding the image block, until all raw image data processing is completed and the process is exited;

[0031] (5) dividing the raw image data into image blocks of equal size, inputting the image blocks into a region matching calculation module in sequence, judging whether the image blocks belong to a hot region, and entering step (6);

[0032] (6) when an image block is detected to belong to a hot region, inputting the image block into a high-precision detection and recognition module for recognition processing and outputting a detection result, otherwise discarding the image block, until all raw image data processing is completed.

[0033] Compared with the prior art, the present application has the following advantages:

[0034] (1) The present application uses a fast screening network to quickly screen suspected target region images in massive raw data according to the sparse characteristics of ocean targets, solving the problem of real-time processing of massive raw data under limited resources;

[0035] (2) The present application constructs a global hot region target information database in the on-board processing system, quickly matches prior fixed targets and targets guided by intelligence information, eliminates a large amount of invalid data, and improves the processing rate of massive raw data;

[0036] (3) The present application uses different computing nodes for different imaging application modes, and through distributed collaborative processing, the processing rate of massive data is solved while the processing accuracy is guaranteed;

[0037] (4) The fast screening network of the present application has a depth of only 12 layers, and has low computational complexity, and can effectively distinguish suspected targets in complex sea clutter interference scenarios such as clouds and islands on orbit. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1This is a schematic diagram of the distributed collaborative fast and high-precision processing system for massive onboard image data of the present invention.

[0039] Figure 2 This is a flowchart of the wide-area search and processing procedure for ocean targets in this invention;

[0040] Figure 3 This is a schematic diagram of the rapid screening network structure of the present invention;

[0041] Figure 4 This is a flowchart illustrating the detailed investigation process for key objectives of this invention.

[0042] Figure 5 This is a schematic diagram of the high-precision detection network structure of the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] A distributed, collaborative, fast, and high-precision system for processing massive amounts of image data on space, such as Figure 1 As shown, it includes: an imaging mode judgment module, a region matching calculation module, a fast screening module, and a high-precision detection and recognition module.

[0045] This invention includes two working modes: wide-area search and processing of ocean targets and detailed investigation and processing of key targets. Both modes first extract images of suspected target areas from the original massive data based on high-speed parallel computing nodes, using a fast filtering network and a spaceborne hotspot target information database. Then, the images of suspected target areas are sent to a high-precision detection and recognition module for high-precision and refined processing. Through distributed collaborative processing of heterogeneous computing nodes, the processing accuracy is guaranteed while significantly reducing the original data rate.

[0046] Specifically, such as Figure 2 As shown, the wide-area search and processing mode for ocean-going targets is mainly used for the rapid wide-area search and discovery of ocean-going ship targets. The imaging mode judgment module inputs the massive amount of raw data from the camera into the fast filtering module. First, the image is processed into blocks through multiple parallel computing nodes, and then the fast filtering network is used to filter the massive amount of high-speed raw data to extract images of suspected target areas, which greatly reduces the raw data rate. Then, the images of suspected target areas are sent to the high-precision detection and recognition module for fine processing. By utilizing the sparsity features of targets, the processing speed is greatly improved while ensuring the processing accuracy.

[0047] exist Figure 2 In the illustrated embodiment, the fast screening module comprises two high-speed parallel computing units composed of FPGA chips, and its specific workflow is as follows:

[0048] The FPGA receives high-speed raw image data and stores the data in a DDR for caching; reads an image block I with a size of MxN from the DDR; and determines whether there is a suspicious target in the image block I by using a fast screening network.

[0049] Further, the fast screening network is a convolutional neural network built using 1x1 and 3x3 convolution kernels, and when performing convolution operations, all convolution steps are 1, thereby ensuring that each channel of the input and output data of the convolution layer is consistent in the size dimension, that is, the number of output data channels is the same as the number of convolution kernels; and the feature map output after convolution is activated by using a Relu activation function. As shown in Figure 3 When the remote sensing image is input into the network, 32 1x1 convolution kernels are first used for convolution, and then 64 3x3 convolution kernels are used for processing. The feature map obtained by the second layer of convolution is input into a multi-channel feature extraction module for processing, the module has three channels, the first channel is connected with 64 3x3 convolution kernels after 32 1x1 convolution kernels, the second channel is connected with 64 1x1 convolution kernels after 32 3x3 convolution kernels, and the third channel is 64 1x1 convolution kernels, and each multi-channel feature extraction module is connected with a Max pooling layer and a convolution layer composed of 64 3x3 convolution kernels, to form a feature extraction unit. The Max pooling layer is a maximum pooling layer. After passing through three feature extraction units, the obtained feature map is processed by using a convolution layer including 32 3x3 convolution kernels, and the processing result is input into a fully connected layer and a Softmax layer to obtain the final classification result. The fully connected layer is used to summarize all results, and the Softmax layer is used to normalize the results and output the final classification result. The fast screening network has only 12 layers in depth, has low computational complexity, and can realize effective differentiation of suspected targets in complex sea clutter interference scenarios such as clouds, fog and small islands on orbit.

[0050] As shown in Figure 4 The key target detailed investigation processing mode is mainly used for detailed investigation of key targets such as ports and airports, and in the case of ground human intelligence or other satellite (such as electronic satellite) guidance information. The imaging mode judgment module inputs the massive raw data input by the camera into the area matching calculation module. First, the raw data is processed in parallel by multiple parallel computing nodes, the latitude and longitude information corresponding to each image block is calculated, and the area matching calculation is realized according to the prior information such as the on-board hot target information library. After the suspected target area image is extracted, it is sent to the high-precision detection and recognition module for fine processing. By using the fixed area target prior information, the processing accuracy is ensured while the processing speed is greatly reduced.

[0051] Specifically, first, the earth surface is divided into equal size grids according to the latitude and longitude index, and the size of each grid is defined as mxn. Through grid division of the earth surface, the range of the hot spot area of any shape can be accurately described by combining different grids. When the data is filtered in the on-board processing system, the corresponding target area image can be accurately extracted, irrelevant data processing is reduced, the processing rate and accuracy are improved, and the user's demand for detailed observation of any interested area is met. The selection of the size of mxn is mainly determined by the range of the ground area corresponding to the image block size of the on-board processing system. The closer the size of mxn is to the image block size, the more accurate the extracted suspected area image block and the more accurate the hot spot area description. Therefore, in actual engineering, the size of mxn needs to be selected by comprehensively considering the image block size of the on-board processing system and the shape rules of different hot spot areas.

[0052] After the size of mxn is selected, each grid on the earth surface is marked and numbered, denoted as Z k (i,j).

[0053] Wherein, -floor(90 / m)≤i≤floor(90 / m),

[0054] -floor(90 / n)≤j≤floor(90 / n),

[0055] 1≤k≤(180*2*90*2) / (m*n).

[0056] A certain hot spot area A x can be composed of a plurality of basic grids Z k (i,j):

[0057] A x ={Z x ,1≤x≤(180*2*90*2) / (m*n)}.

[0058] The global hot spot area information library T is:

[0059] T=A x ∪B y ∪C z …

[0060] Wherein, A x , B y , C z represent the key sea area, port, airport and other hot spot area sets respectively.

[0061] The global hot spot area information library is first constructed on the ground and then uploaded to the on-board processing system. During the satellite operation, it can be dynamically updated, maintained and upgraded online according to the demand.

[0062] In Figure 4 In the embodiment shown, the rapid screening module includes a high-speed parallel computing unit composed of an FPGA chip, and the specific working process is as follows:

[0063] S1: The FPGA receives high-speed raw image data and stores it in DDR for caching;

[0064] S2: Read an image block I of size MxN from DDR;

[0065] S3: Calculate the ground longitude and latitude coordinates I(k, l) corresponding to the center point of the current image using the GPS, attitude and other information in the image auxiliary data, and perform grid normalization processing I(i, j); where i=floor(180 / M), j=floor(90 / N);

[0066] S4: Match I(i, j) with the hotspot area information library T, when I(i, j)T, send the image block I to the high-precision detection and recognition module for high-precision detection and recognition processing, otherwise discard the image block;

[0067] Repeat S2-S4, empty DDR after processing the image data in DDR, and read the next scene image data.

[0068] The high-precision detection and recognition module of the application meets the demand for high-precision detection and recognition of complex background targets under the condition of limited resources on the satellite. The high-precision detection network based on super-resolution auxiliary branch is used to process the input image. The network takes YOLOv5s structure as the baseline, and introduces multi-modal data fusion structure, high-resolution information preservation network and auxiliary super branch to ensure high-precision and low-false-alarm detection and recognition of targets in complex background. The super-resolution branch does not participate in the inference stage, which does not introduce additional calculation for the detection network and does not affect the inference speed, and is easy to deploy on the resource-limited satellite, so as to achieve the optimization of detection accuracy and processing rate,

[0069] Specifically, as shown in Figure 5 The high-precision detection network includes a multi-modal data fusion unit, a high-resolution information preservation unit, a backbone network and an auxiliary super-resolution learning branch.

[0070] The multi-modal data fusion unit: normalizes the multi-modal images with different resolutions in the original image data to the same resolution, forms a fusion image through multi-source data fusion and inputs it to the high-resolution information preservation unit; the multi-modal images include panchromatic images, multispectral images and near-infrared images.

[0071] Specifically, the complementarity between different modalities of visible light panchromatic, multispectral, and infrared is utilized to design a detection structure based on multi-modal adaptive fusion, so as to obtain more information to distinguish targets and further improve the detection accuracy and universality of the algorithm. First, the input panchromatic image, RGB image, and near-infrared image are normalized to the [0, 1] interval, and they are connected with relatively low computational complexity to speed up the inference. Specifically, the fused image is defined as:

[0072] X = Concat(Q, R, G, B, I)

[0073] where the fused image is C represents the number of channels, H and W represent the height and width of the image, respectively, {Q}, {R, G, B}, and {I} represent the panchromatic image, RGB image, and near-infrared image, respectively, and Concat(·) represents the connection operation along the channel dimension. Then, X is down-sampled to 1 / n of the original image size to complete the super-resolution module and accelerate the training process. The sampled image is represented as generated by: X' = D(X), where D(·) represents an n-time down-sampling operation using bilinear interpolation. Then, the down-sampled result is input into the backbone to produce multi-level features.

[0074] High-resolution information preservation unit: details of the fused image are preserved and enhanced to obtain a high-resolution image with enhanced details and input into the backbone network.

[0075] Specifically, a high-resolution information preservation network is constructed to preserve the high-resolution characteristics of remote sensing targets and overcome the problem of missing small target information. The Focus module in the YOLOv5 backbone divides the image by intervals in the spatial domain, and then reorganizes the new image to adjust the size of the input image. Specifically, this operation collects a value for each pixel in the image, and then reconstructs it to obtain a smaller complementary image. The size of the reconstructed image decreases with the increase of the number of channels. Therefore, it will cause the resolution of small targets to decrease and the spatial information to be lost. Considering that the detection of small targets relies more on higher resolution, the Focus module is abandoned and a specific convolution operation is used instead to prevent the loss of small targets caused by resolution reduction.

[0076] Backbone network: the high-resolution image with enhanced details is extracted to obtain low-level and high-level features of the image, and a high-precision detection and recognition model is used to complete the judgment of the suspicious target class in the image and calculate its position and confidence.

[0077] The auxiliary super-resolution learning branch: the low-level and high-level features of the image obtained by the backbone network are used for super-resolution learning when the ground system trains the high-precision detection and recognition model, and the local texture, detail features and high-level semantic information of the complex background image are obtained to modify and improve the high-precision detection and recognition model of the backbone network. The auxiliary super-resolution learning branch includes an encoder and a decoder. The encoder matches the low-level features of the image to the spatial size of the high-level features, and then combines the low-level features with the high-level features. The decoder decodes the combined information of the low-level features and the high-level features, and outputs super-resolution features.

[0078] Specifically, the auxiliary super-resolution learning branch is a unit for the ground system to train a high-precision detection and recognition model of the backbone network. The auxiliary super-resolution learning branch does not participate in on-board reasoning calculation. By designing a learning branch based on auxiliary super-resolution, the network's high-resolution information retention capability is improved, and the detection accuracy of the algorithm is further improved. The feature size reserved in the backbone network for multi-scale detection is much smaller than the original input image. Existing methods mostly perform upsampling operations to restore the feature size. Unfortunately, due to the loss of information in texture and pattern, the effect of this method is limited. Therefore, it is not suitable to use this method to detect small targets in remote sensing images. To solve this problem, the present method introduces an auxiliary super-resolution branch. Specifically, the super-resolution structure can be regarded as a simple encoder-decoder model. The low-level and high-level features of the backbone network are selected as the input of the super-resolution branch to obtain local texture, pattern and semantic information. In the encoder, the low-level features are first subjected to a CBR module and an upsampling operation to match the spatial size of the high-level features, and then the low-level and high-level features are combined using a concatenation operation and two CBRD modules. The CBR module includes convolution, batch normalization and ReLU activation function, while the CBRD module includes an additional drop-out. For the decoder, the low-resolution features are upgraded to the high-resolution space, where the output size of the super-resolution module is twice that of the input image. The decoder is implemented using three deconvolution layers. The super-resolution branch guides the relevant learning of the spatial dimension and transfers it to the main branch, thereby improving the performance of target detection. The super-resolution branch is completely removed during the reasoning stage and does not participate in the detection reasoning stage, which does not introduce additional computational load for the detection network and does not affect the reasoning speed. Moreover, since the input image is a relatively low-resolution image, compared with ordinary detection networks, the effect of acceleration can be achieved to meet the demand for efficient detection tasks under the condition of limited on-board resources.

[0079] The method for distributed collaborative fast high-precision processing of massive on-board image data comprises the following steps:

[0080] 1) receiving the original image data input by the on-board camera;

[0081] 2) judging the imaging mode of the satellite, when the satellite works in the wide-area search imaging mode of the ocean target, entering step 3), when the satellite works in the detailed imaging mode of the point target, entering step 5);

[0082] 3) dividing the original image data into image blocks of equal size and inputting the image blocks into the fast screening module in sequence for suspected target screening processing, entering step 4);

[0083] 4) when detecting that the image block has a suspected target, inputting the image block into the high-precision detection and recognition module for recognition processing and outputting the detection result, otherwise discarding the image block, until the processing of all original image data is completed, exiting the method;

[0084] 5) dividing the original image data into image blocks of equal size, inputting the image blocks into the region matching calculation module in sequence, calculating the ground longitude and latitude coordinates corresponding to the center points of the image blocks by using the GPS, attitude and other information in the image auxiliary data, determining the basic grid number where the center points of the image blocks are located according to the longitude and latitude coordinates, judging whether the basic grid belongs to the hot region, entering step 6);

[0085] 6) when detecting that the image block belongs to the hot region, inputting the image block into the high-precision detection and recognition module for recognition processing and outputting the detection result, otherwise discarding the image block, until the processing of all original image data is completed, exiting the method.

[0086] The contents not described in detail in the specification of the present application are the known technology of the person skilled in the art.

Claims

1. A distributed, collaborative, fast, and high-precision system for processing massive amounts of spaceborne image data, characterized in that: include: The module comprises an imaging mode determination module, a region matching calculation module, a fast filtering module, and a high-precision detection and recognition module, among which: Imaging mode determination module: Determines the current imaging mode of the remote sensing satellite. When the remote sensing satellite is operating in the wide-area search imaging mode for ocean targets, the original image data is input into the fast filtering module. When the remote sensing satellite is operating in the detailed investigation imaging mode for point targets, the original image data is input into the region matching calculation module. The fast filtering module filters out images suspected of containing suspicious targets from the raw image data and inputs them into the high-precision detection and recognition module, discarding the remaining images; Region matching calculation module: Constructs hotspot regions, filters out images belonging to the hotspot regions from the original image data and inputs them into the high-precision detection and recognition module, discarding the remaining images; High-precision detection and recognition module: performs target detection and recognition on the input image and outputs the detection results; the high-precision detection and recognition module includes: a multimodal data fusion unit, a high-resolution information preservation unit, a backbone network, and an auxiliary super-resolution learning branch, wherein: Multimodal data fusion unit: unifies the original images to the same resolution, forms a fused image through multi-source data fusion, and inputs it to the high-resolution information preservation unit; the original images include panchromatic images, multispectral images, and near-infrared images; High-resolution information preservation unit: performs detail preservation and enhancement on the fused image to obtain a high-resolution image with enhanced details, which is then input into the backbone network; Backbone network: Performs hierarchical feature extraction on high-resolution images with enhanced details to obtain low-level and high-level features of the image. Uses a high-precision detection and recognition model to determine the category of suspicious targets in the image and calculates the location and confidence level of the suspicious targets. The auxiliary super-resolution learning branch utilizes the low-level and high-level features of the image obtained from the backbone network for super-resolution learning. After acquiring local texture, detail features, and high-level semantic information of complex background images, it corrects the high-precision detection and recognition model of the backbone network.

2. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 1, characterized in that, The fast filtering module includes one or more data processing units. Each data processing unit uses a fast filtering network to filter the input image. The fast filtering network includes a first convolutional layer, a second convolutional layer, at least one feature extraction unit, a third convolutional layer, a fully connected layer, and a softmax layer arranged sequentially. The first convolutional layer includes 32 layers. The second convolutional layer contains 64 convolutional kernels. The convolutional kernel is used, and the feature extraction unit includes a multi-channel feature extraction unit, a max pooling layer, and a fourth convolutional layer arranged sequentially. The multi-channel feature extraction unit includes a first channel, a second channel, and a third channel, with the first channel having 32 channels. 64 convolutional kernels are then connected. The convolution kernel has 32 kernels in the second channel. 64 convolutional kernels are then connected. The convolution kernel, the third channel has 64 kernels. The fourth convolutional layer has 64 convolutional kernels. The third convolutional layer comprises 32 convolutional kernels. The convolution kernels; the feature map output after convolution by the convolution kernels in the first convolutional layer, the second convolutional layer, the first channel, the second channel, the third channel, the third convolutional layer, or the fourth convolutional layer is activated using the ReLU activation function.

3. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 2, characterized in that, The convolution kernel and The convolution kernels have a step size of 1.

4. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 2, characterized in that, The number of feature extraction units is 3 to 5, which are connected in series.

5. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 1, characterized in that, The auxiliary super-resolution learning branch includes an encoder and a decoder. The encoder matches the low-level features of the image to the spatial size of the high-level features, and then merges the low-level features with the high-level features. The decoder decodes the information after merging the low-level features and the high-level features and outputs the super-resolution features.

6. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 1, characterized in that, The region matching calculation module constructs hotspot regions, including: The Earth's surface is divided into equal-sized basic grids according to latitude and longitude indexes. Each basic grid is marked and numbered. Based on the information marked on the ground, one or more hotspot regions are formed, and each hotspot region consists of one or more basic grids.

7. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 6, characterized in that, The region matching calculation module extracts images belonging to hotspot regions from the original image, including: The original image is divided into image blocks of equal size. The ground latitude and longitude coordinates corresponding to the center point of the image block are calculated using GPS and attitude information from the image auxiliary data. The basic grid number of the center point of the image block is determined based on the latitude and longitude coordinates. Then, it is determined whether the basic grid belongs to a hotspot area. If it does, the image block is extracted; otherwise, it is discarded.

8. The spaceborne massive image data distributed collaborative fast high-precision processing system according to claim 1, characterized in that, The fast filtering module and the region filtering module include multiple parallel data processing units.

9. An image processing method using the distributed collaborative fast and high-precision processing system for massive spaceborne image data as described in any one of claims 1 to 8, characterized in that, Includes the following steps: (1) Receive raw image data input from the spaceborne camera; (2) Determine the imaging mode of the satellite. When the satellite is working in the wide-area search imaging mode for ocean targets, proceed to step (3). When the satellite is working in the detailed investigation imaging mode for point targets, proceed to step (5). (3) Divide the original image data into image blocks of equal size and input them into the fast filtering module in sequence for suspected target filtering processing, and proceed to step (4). (4) When a suspected target is detected in an image block, the image block is input into the high-precision detection and recognition module for recognition processing and the detection result is output. Otherwise, the image block is discarded until all the original image data processing is completed and the process is exited. (5) Divide the original image data into image blocks of equal size, input the image blocks into the region matching calculation module in turn, determine whether the image block belongs to the hot spot region, and proceed to step (6). (6) When an image block is detected to be a hotspot area, the image block is input into the high-precision detection and recognition module for recognition processing and the detection result is output. Otherwise, the image block is discarded until all original image data processing is completed.

Images