Time-space domain combined space detection image dark target detection method and system

Abstract

The invention provides a time-space domain combined space detection image dark and weak target detection method and system. The method comprises the following steps: slicing an original space target detection image; key points are extracted from each image slice; constructing topological space description features of the key points, and performing inter-frame registration of adjacent frame images; calculating the inter-frame difference of adjacent frames after registration; the inter-frame difference results are superposed; space target training and detection are carried out through a Yolov5 deep learning network; and screening and combining the detection results, and outputting a final detection result. According to the method, the fixed star residual error after inter-frame difference is effectively reduced under the condition of a relatively large detection field of view, the detection performance of a detected image space target is improved, and the method has universality for working modes of staring targets and staring fixed stars in space detection.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and specifically, relates to a method and system for detecting dim targets in space detection images in a combined time-space domain, especially a method for detecting moving dim targets in high-sensitivity space-based optical space detection images under a dense star background, based on a combined time-space domain. Background Art

[0002] With the increasing frequency of human space activities, the number of artificial satellites and space debris is increasing year by year. The detection of space targets is an important means to ensure the normal operation of spacecraft on-orbit missions. Compared with ground-based detection, space-based optical observation has the advantage of not being affected by the atmosphere. However, in space detection images, space targets mainly have point target characteristics with low brightness, few pixels, and no texture information. In addition, the characteristics of stars and space targets are similar, which increases the difficulty of detecting space targets.

[0003] The existing space target detection methods mainly include image difference method and detection method based on motion information. The image difference method extracts star points through morphological processing, uses grayscale and other features to match star points, sets a threshold after inter-frame difference to determine whether it is a space target; the method based on motion information models the inter-frame motion of the target, estimates the target displacement, and realizes inter-frame compensation. However, both methods have some problems, such as: the image field of view used for experimental verification is small and the stars are relatively sparse, so the method has poor detection effect on dense star background; the inter-frame registration accuracy directly affects the detection results, and deviations in registration will cause the star background to be unable to be completely subtracted, forming a crescent-shaped residual and causing false detection; the method based on deep learning usually relies on the morphological texture and other features of the target for detection, but due to the similar characteristics of stars and space targets, it cannot achieve a good detection effect.

[0004] The patent document "Key frame screening method based on distribution of targets of interest" (CN113112519A) combines traditional technology and deep learning technology to extract the underlying features and deep features of the image, and performs multi-feature fusion to improve the robustness of the extracted key frames. However, due to its relatively simple application background, it does not solve the problem that the target is easily submerged by background clutter and noise, making detection difficult.

[0005] The patent document "A method and system for dynamic gesture recognition based on gesture key points" (CN109190461A) extracts gesture key points from the image and uses the key points to obtain the spatiotemporal information of the dynamic gesture. However, it is a feature modeling strategy based on spatiotemporal structural statistics and is a non-parametric learning technology. It still does not solve the problem of false alarms for background residuals and tracking multiple trajectories for a single target. Summary of the invention

[0006] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for detecting dim targets in spatial detection images by combining time and space domains.

[0007] A method for detecting dim targets in spatial detection images in a combined temporal and spatial domains according to the present invention includes:

[0008] Step S1: Slice the original space target detection image with equal step length and overlapping areas to obtain image slices to be detected after cutting;

[0009] Step S2: extract key points for each image slice;

[0010] Step S3: constructing the topological space description features of the key points, and performing inter-frame registration of adjacent frame images according to the key point extraction results of adjacent frames;

[0011] Step S4: Calculate the inter-frame difference of adjacent frames after registration to obtain the inter-frame difference result;

[0012] Step S5: superimposing the inter-frame difference results to obtain a background suppression result;

[0013] Step S6: Perform spatial target training and detection on the background suppression result through the Yolov5 deep learning network to obtain the detection result;

[0014] Step S7: Screen and merge the test results and output the final test results.

[0015] Preferably, step S2 comprises:

[0016] Step S2.1: binarizing the image slice to obtain a binary image;

[0017] Step S2.2: Calculate the connected domain of the stars in the binary image, where the connected domain is an eight-neighborhood domain;

[0018] Step S2.3: Calculate the centroid of the connected domain by grayscale weighted calculation, according to The obtained centroid is taken as the key point;

[0019] Among them, y′ represents the coordinates of the centroid row vector;

[0020] z′ represents the centroid column vector coordinates;

[0021] y″ represents the row coordinate of the i-th pixel in the connected domain;

[0022] z″ represents the column coordinates of the i-th pixel in the connected domain;

[0023] h represents the gray value of the i-th pixel in the connected domain;

[0024] N represents a total of N pixels in the connected domain.

[0025] The step S3 comprises:

[0026] Step S3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points;

[0027] Step S3.2: performing key point matching to obtain matching point pairs;

[0028] The matching is to calculate the feature similarity of key points in the current frame and adjacent frames.

[0029] Step S3.3: aligning adjacent frames to the starry sky background based on the current frame according to calculation;

[0030] The calculation is to obtain the transformation matrix from the adjacent frame image to the current frame image based on the matched point pairs based on homography transformation and estimation;

[0031] The homography transformation and estimation formula are:

[0032] in, represents the homography matrix;

[0033] h ij represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3;

[0034] x represents the horizontal coordinate of the key point in the adjacent frame image;

[0035] y represents the vertical coordinate of the key point in the adjacent frame image;

[0036] x′ is the horizontal coordinate of the same key point in the current frame;

[0037] y′ represents the ordinate of the corresponding key point in the current frame.

[0038] The step S6 comprises:

[0039] Step S6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters;

[0040] Step S6.2: Detect the space target detection image using the trained network parameters to obtain a candidate detection frame;

[0041] The Yolov5 deep learning network is a YOLOv5s model.

[0042] The step S7 comprises:

[0043] Step S7.1: sorting the obtained multiple candidate detection boxes according to the confidence scores;

[0044] Step S7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them. The formula is:

[0045] Among them, IoU represents the degree of overlap;

[0046] A and B represent the candidate detection box with the highest score and the traversed candidate detection box in the overlapping area of ​​the image slices, respectively;

[0047] Step S7.3: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than the threshold, the currently traversed candidate detection frame is deleted; when the intersection-and-union ratio is less than the threshold, the candidate detection frame is retained;

[0048] Repeat steps S7.1 to S7.3 until all candidate detection frames have been traversed.

[0049] Preferably, the step S3.1 comprises:

[0050] Step S3.1.1: Take any key point p i As the center of the circle, a set number of concentric circles are constructed around it and divided equally into a set number of regions;

[0051] Step S3.1.2: For key point p i The key points around j , p j The coordinates in the Cartesian coordinate system are transformed into the key point p i is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j , the corresponding area number is q j for:

[0052] 0≤θ j ≤2π,0≤ρ j ≤R;

[0053] Where R represents the maximum concentric circle radius;

[0054] m represents the unit radian when quantizing the polar angle;

[0055] n is the unit distance when quantizing the polar diameter;

[0056] Step S3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature Fi .

[0057] The step S3.2 comprises:

[0058] Step S3.2.1: Set the current frame I n The key point p i The position in the Cartesian coordinate system (x i ,y i ) as the search starting point, in the adjacent frame I n +Δ searches for a key point p near the starting point k And calculate its position feature F k ;

[0059] Step S3.2.2: Calculate the distance between the position features and the grayscale difference of the key points in the original image slice;

[0060] Step S3.2.3: If the formula condition is met, the current frame I n The key point p i and adjacent frame I n +Δ key point p k If the match is successful, it will be regarded as a pair of matched points; if it is not satisfied, the calculation will continue;

[0061] The formula conditions are

[0062] Among them, q represents the region number in the feature;

[0063] f(p i ) represents the key point p i Grayscale value on the original image;

[0064] f(p k ) represents the key point p k Grayscale value on the original image;

[0065] th1 is the threshold of the distance between features;

[0066] th2 is the grayscale difference threshold;

[0067] Traverse all key points and execute steps S3.2.1 to S3.2.3 to obtain all matching point pairs.

[0068] Preferably, the formula calculated in step S4 is:

[0069]

[0070] Among them, Sat_Value represents the saturation value of the image;

[0071] D N(i, j) represents the difference result between adjacent frames at position (i, j);

[0072] f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame;

[0073] f N-1 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame is registered.

[0074] In step S5, the current frame is registered and differentiated with multiple adjacent frames, and the registration difference result obtained after the registration difference between the jth frame and the kth frame is D(I j ,I j-k ), by accumulating n frames between frames, taking the average value of n frame registration difference results, the formula is

[0075] Among them, I j represents the spatial target signal in the jth frame;

[0076] I j-1 Represents the spatial target signal from the jth frame to the kth frame;

[0077] I j ' represents the enhanced spatial target signal in the jth frame.

[0078] According to the present invention, a spatial detection image dim target detection system combined with time and space domains includes:

[0079] Module M1: Slice the original space target detection image with equal steps and overlapping areas to obtain the image slices to be detected after cutting;

[0080] Module M2: Extract key points for each image slice;

[0081] Module M3: construct the topological space description features of key points, and perform inter-frame registration of adjacent frame images based on the key point extraction results of adjacent frames;

[0082] Module M4: Calculate the inter-frame difference of adjacent frames after registration to obtain the inter-frame difference result;

[0083] Module M5: superimpose the inter-frame difference results to obtain the background suppression result;

[0084] Module M6: Perform spatial target training and detection on the background suppression results based on the Yolov5 deep learning network to obtain the detection results;

[0085] Module M7: Filter and merge the test results and output the final test results.

[0086] Preferably, the module M2 comprises:

[0087] Module M2.1: Binarize the image slices to obtain a binary image;

[0088] Module M2.2: Calculate the connected domain of stars in the binary image, where the connected domain is an eight-neighborhood domain;

[0089] Module M2.3: Find the centroid of the connected domain by grayscale weighted calculation, according to

[0090] The obtained centroid is taken as the key point;

[0091] Among them, y′ represents the coordinates of the centroid row vector;

[0092] z′ represents the centroid column vector coordinates;

[0093] y″ represents the row coordinate of the i-th pixel in the connected domain;

[0094] z″ represents the column coordinates of the i-th pixel in the connected domain;

[0095] h represents the gray value of the i-th pixel in the connected domain;

[0096] N represents a total of N pixels in the connected domain.

[0097] The module M3 comprises:

[0098] Module M3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points;

[0099] Module M3.2: Perform key point matching to obtain matching point pairs;

[0100] The matching is to calculate the feature similarity of key points in the current frame and the adjacent frames;

[0101] Module M3.3: align adjacent frames to the starry sky background based on the current frame according to calculation;

[0102] The calculation is to obtain the transformation matrix from the adjacent frame image to the current frame image based on the matched point pairs based on homography transformation and estimation;

[0103] The homography transformation and estimation formula are:

[0104] in, represents the homography matrix;

[0105] h ij represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3;

[0106] x represents the horizontal coordinate of the key point in the adjacent frame image;

[0107] y represents the vertical coordinate of the key point in the adjacent frame image;

[0108] x′ is the horizontal coordinate of the same key point in the current frame;

[0109] y′ represents the ordinate of the corresponding key point in the current frame.

[0110] The module M6 comprises:

[0111] Module M6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters;

[0112] Module M6.2: Use the trained network parameters to detect the space target detection image and obtain the candidate detection box;

[0113] The Yolov5 deep learning network is a YOLOv5s model.

[0114] The module M7 comprises:

[0115] Module M7.1: Sort the obtained multiple candidate detection boxes according to the confidence scores;

[0116] Module M7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them. The formula is

[0117] Among them, IoU represents the degree of overlap;

[0118] A and B represent the candidate detection box with the highest score and the traversed candidate detection box in the overlapping area of ​​the image slices, respectively;

[0119] Module M7.3: Judgment: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than the threshold, the currently traversed candidate detection frame is deleted; when the intersection-and-union ratio is less than the threshold, the candidate detection frame is retained;

[0120] Modules M7.1 to M7.3 are triggered repeatedly until all candidate detection frames are traversed and the process stops.

[0121] Preferably, the module M3.1 comprises:

[0122] Module M3.1.1: Take any key point p i As the center of the circle, a set number of concentric circles are constructed around it and divided equally into a set number of regions;

[0123] Module M3.1.2: For key point p iThe key points around j , p j The coordinates in the Cartesian coordinate system are transformed into the key point p i is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j , the corresponding area number is q j for:

[0124] 0≤θ j ≤2π,0≤ρ j ≤R;

[0125] Where R represents the maximum concentric circle radius;

[0126] m represents the unit radian when quantizing the polar angle;

[0127] n is the unit distance when quantizing the polar diameter;

[0128] Module M3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature F i .

[0129] The module M3.2 includes:

[0130] Module M3.2.1: Set the current frame I n The key point p i The position in the Cartesian coordinate system (x i ,y i ) as the search starting point, in the adjacent frame I n +Δ searches for a key point p near the starting point k And calculate its position feature F k ;

[0131] Module M3.2.2: Calculate the distance between position features and the grayscale difference of key points in the original image slice;

[0132] Module M3.2.3: If the formula conditions are met, the current frame I n The key point p i and adjacent frame I n +Δ key point p k If the match is successful, it will be regarded as a pair of matched points; if it is not satisfied, the calculation will continue;

[0133] The formula conditions are

[0134] Among them, q represents the region number in the feature;

[0135] f(p i ) represents the key point p i Grayscale value on the original image;

[0136] f(p k ) represents the key point p k Grayscale value on the original image;

[0137] th1 is the threshold of the distance between features;

[0138] th2 is the grayscale difference threshold;

[0139] Traverse all key points and trigger modules M3.2.1 to M3.2.3 to obtain all matching point pairs.

[0140] Preferably, the formula calculated in the module M4 is:

[0141]

[0142] Among them, Sat_Value represents the saturation value of the image;

[0143] D N (i, j) represents the difference result between adjacent frames at position (i, j);

[0144] f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame;

[0145] f N-2 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame is registered.

[0146] In the module M5, the current frame is registered and differentiated with multiple adjacent frames, and the registration difference result obtained after the registration difference from the jth frame to the kth frame is D(I j ,I j-k ), by accumulating n frames between frames, taking the average value of n frame registration difference results, the formula is

[0147] Among them, I j represents the spatial target signal in the jth frame;

[0148] I j-1 Represents the spatial target signal from the jth frame to the kth frame;

[0149] I j ' represents the enhanced spatial target signal in the jth frame.

[0150] According to a computer-readable storage medium storing a computer program provided by the present invention, when the computer program is executed by a processor, the steps of the method for detecting dim targets in spatial detection images in a combined time-space domain are implemented.

[0151] An electronic device provided according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the method for detecting dim targets in spatial detection images by combining time and space domains are implemented.

[0152] Compared with the prior art, the present invention has the following beneficial effects:

[0153] 1. The present invention detects objects in image space under a larger detection field of view by performing detection after image segmentation and then merging the detection results using non-maximum suppression.

[0154] 2. The topological space description feature extraction method with scale and rotation invariance provided by the present invention realizes sub-pixel inter-frame registration of images, effectively reduces the star residual after inter-frame difference, thereby improving the detection performance.

[0155] 3. The present invention combines the traditional inter-frame registration difference method with the deep learning intelligent detection algorithm, which not only solves the problem that the traditional threshold method is prone to false alarms for star residuals, but also solves the problem that directly using the intelligent detection algorithm cannot distinguish between stars and space targets which are both point targets.

[0156] 4. The detection method provided by the present invention does not rely on the point or line features of stars and space targets for detection, and does not require switching different methods and parameters according to different detection modes. It is universal for the working modes of "staring at targets" and "staring at stars" in space detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0157] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0158] Figure 1 The flowchart of the method for detecting dim targets in spatial detection images jointly in the spatiotemporal and temporal domains is shown in FIG.

[0159] Figure 2 Schematic diagram of segmentation of overlapping areas of spatial detection images.

[0160] Figure 3 Schematic diagram of key point extraction results for adjacent frames.

[0161] Figure 4 Schematic diagram of region division for constructing topological space description features.

[0162] Figure 5 Schematic diagram for comparison of registration difference results.

[0163] Figure 6 Schematic diagram of repeated detection boxes detected for the same spatial target in adjacent detection image slices. DETAILED DESCRIPTION

[0164] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0165] In complex space environments, due to the similar imaging characteristics of stellar targets and space targets, intelligent algorithms based on deep learning models cannot distinguish stellar targets from space targets. Traditional detection algorithms have the disadvantages of low accuracy and high false alarm rate. In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for detecting dark targets in space detection images in a joint space-time domain.

[0166] In the background of dense stars (as shown in Table 1), faint targets are easily submerged by background clutter and noise, and the similar characteristics of stars and space targets make it difficult to detect faint space targets. By designing features with scale and rotation invariance for inter-frame matching and combining them with deep learning methods, the detection of faint space targets in space detection images with high sensitivity (21 magnitude) detection cameras within a large field of view is achieved. It can adapt to different imaging modes of space detection such as "staring at targets" and "staring at stars", and improve the performance and robustness of the detection method.

[0167] Table 1. Relationship between the number of stars of different magnitudes per square degree and the galactic latitude

[0168]

[0169] According to the present invention, a method for detecting dim targets in spatial detection images in a combined time-space domain is provided. Figure 1 As shown, the following steps are included:

[0170] Step S1: Slice the original space target detection image with equal step length and overlapping areas. Figure 2 For example, the step size and overlapping area can be preset to obtain multiple cropped image slices to be detected.

[0171] Step S2: extract key feature points for each image slice;

[0172] The image slices are binarized, the connected domains are extracted from the binarized images, and the centroid of the connected domain is calculated. The centroid position is used as the key point to locate the star.

[0173] In more preferred embodiments, step S2 includes:

[0174] Step S2.1: binarize the image slices using the Otsu method to obtain a binary image;

[0175] Step S2.2: Calculate the connected domain of the star in the binary image by using a connected domain extraction algorithm, wherein the connected domain is an eight-neighborhood domain;

[0176] Step S2.3: Find the centroid of each connected domain, use the grayscale weighted method to calculate, and use all the centroids obtained as key points according to the formula:

[0177]

[0178] Among them, y′ represents the coordinates of the centroid row vector;

[0179] z′ represents the centroid column vector coordinates;

[0180] y″ represents the row coordinate of the i-th pixel in the connected domain;

[0181] z″ represents the column coordinates of the i-th pixel in the connected domain;

[0182] h represents the gray value of the i-th pixel in the connected domain;

[0183] N represents a total of N pixels in the connected domain.

[0184] Step S3: Describe the features by constructing a topological space with scale and rotation invariance. Figure 3 For example, based on the key point extraction results of adjacent frames, the similarity of key point features is used to achieve inter-frame registration of adjacent frame images.

[0185] In more preferred embodiments, step S3 includes:

[0186] Step S3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points.

[0187] The step S3.1 comprises:

[0188] Step S3.1.1: Figure 4 For example, take any key point p i With the center as the circle, multiple groups of concentric circles are constructed around it and divided equally into multiple areas.

[0189] Step S3.1.2: For key point p i The key points aroundj , p j The coordinates in the Cartesian coordinate system are transformed into the key point p i is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j The corresponding area number is q j for:

[0190] 0≤θ j ≤2π,0≤ρ j ≤R;

[0191] Where R represents the maximum concentric circle radius;

[0192] m represents the unit radian when quantizing the polar angle;

[0193] n is the unit distance when quantizing the polar diameter.

[0194] Step S3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature F i .

[0195] Step S3.2: Find all matching key points in the image slices of the current frame and the adjacent frames, including:

[0196] Calculate the feature similarity of key points in the current frame and adjacent frames to achieve key point matching.

[0197] Step S3.2.1: Set the key point p in the current frame In i The position in the Cartesian coordinate system (x i ,y i ) as the search starting point, and search for the key point p near the starting point in the adjacent frame In+Δ k And calculate its position feature F k .

[0198] Step S3.2.2: Calculate the distance between the position features and the grayscale difference of the key points in the original image.

[0199] Step S3.2.3: If the formula conditions are met, the key point p of the current frame In i and the key point p of the adjacent frame In+Δ k The match is successful as a pair of key points, i.e. point pair. The formula condition is

[0200] Among them, q represents the region number in the feature;

[0201] f(pi ) represents the key point p i Grayscale value on the original image;

[0202] f(p k ) represents the key point p k Grayscale value on the original image;

[0203] th1 is the threshold of the distance between features;

[0204] th2 is the grayscale difference threshold.

[0205] Traverse all key points and perform the above operations to obtain all matching point pairs.

[0206] Step S3.3: Calculate the homography matrix and align adjacent frames to the starry sky background based on the current frame.

[0207] According to the matched point pairs, the transformation matrix from the adjacent frame image to the current frame image is obtained based on homography transformation and estimation, and the adjacent frame is aligned to the current frame.

[0208] The homography matrix formula is:

[0209] in, represents the homography matrix;

[0210] h ij represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3;

[0211] x represents the horizontal coordinate of the key point in the adjacent frame image;

[0212] y represents the vertical coordinate of the key point in the adjacent frame image;

[0213] x′ is the horizontal coordinate of the same key point in the current frame;

[0214] y′ represents the ordinate of the corresponding key point in the current frame.

[0215] In more preferred embodiments, it is assumed that the image coordinates of a point of the imaged rigid body at time k are (x, y) T , at time k+1, the point moves to (x′,y′) T . Just find the optimal homography matrix to get the global motion model parameters. In real detection scenarios, the calculated key point pairs will contain noise, such as the position of the point has a deviation of several pixels, or even the key point pair mismatch. Therefore, in order to make the calculation more accurate, more than 4 point pairs are generally used to calculate the homography matrix. By performing a homography transformation, the adjacent frames are aligned to the starry sky background based on the current frame, so that the positions of most stars in the background can overlap.

[0216] Step S4: Calculate the inter-frame difference results of adjacent frames after registration to achieve star background suppression;

[0217] The inter-frame difference method is used to suppress the star background.

[0218] by Figure 5 For example, the traditional differential method is adaptively improved by fully considering the grayscale characteristics of corresponding pixels between frames. At the same time, the grayscale of the differential result is stretched, so that after the differential of dim targets between frames, the target signal will not be weakened, but the target signal will be greatly enhanced.

[0219] The formula calculated in step S4 is:

[0220]

[0221] Among them, Sat_Value represents the saturation value of the image;

[0222] D N (i, j) represents the difference result between adjacent frames at position (i, j);

[0223] f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame;

[0224] f N-2 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame is registered.

[0225] Figure 5 is the comparison chart of the registration difference results, where Figure 5 (a) is the detection image slice, Figure 5 (b) is the image difference result using ORB registration algorithm. Figure 5 (c) is the result of image registration and difference using the topological space description feature provided by the present invention. From the comparison results, it can be seen that the difference effect is significantly improved after the image registration using the topological space description feature provided by the present invention.

[0226] Step S5: superimposing the difference results of multiple frames to achieve target signal-to-noise ratio enhancement and obtain background suppression results;

[0227] Based on the results of sub-pixel inter-frame registration and inter-frame difference background suppression, the target signal-to-noise ratio is enhanced by the inter-frame accumulation method.

[0228] In more preferred examples, the target is faint and is likely to be submerged by the stellar differential residual. To enhance the signal-to-noise ratio of the faint target, the current frame is registered and differentiated with multiple adjacent frames, and the differential results are superimposed.j ,I j-k );

[0229] Among them, I j represents the spatial target signal in the jth frame;

[0230] I j-1 Represents the spatial target signal from the jth frame to the kth frame;

[0231] The positions of space targets are all related to I j The target positions in are the same, while the distribution of the residual noise and star brightness residuals from the jth frame to the kth frame is random.

[0232] Therefore, by accumulating n frames between frames and taking the average of the n-frame registration difference results, the spatial target signal in the j-th frame is enhanced. The formula is:

[0233]

[0234] Among them, I j ' represents the enhanced spatial target signal in the jth frame.

[0235] Step S6: Using Yolov5 deep learning network to perform spatial target training and detection based on the background suppression results;

[0236] In more preferred embodiments, step S6 includes:

[0237] Step S6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters;

[0238] Step S6.2: Use the trained network parameters to detect the space target detection image.

[0239] The Yolov5 deep learning network is specifically a YOLOv5s model;

[0240] Step S7: filter and merge the detection results of the image slices, and output the final detection results of the original image size.

[0241] The detection results of overlapping areas of image slices are screened and merged through the non-maximum suppression method, and the sub-image detection results are merged to restore the final detection results of the original image size.

[0242] In more preferred embodiments, Figure 6For example, in the case where adjacent image slices with overlapping areas may contain the same target with different center point positions, widths and heights, in order to ensure that the detection frame obtained by the spatial target in the final detection result is unique and to prevent multiple tracks from being tracked for a single spatial target during subsequent track association, the non-maximum suppression technology is used to screen the similar detection frames in the overlapping area, merge the detection results, and retain only the valid detection frames. The step S7 includes:

[0243] Step S7.1: sorting the obtained multiple candidate detection boxes according to their confidence scores;

[0244] Step S7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them according to the formula:

[0245]

[0246] Among them, IoU is the overlap;

[0247] A and B are the candidate detection box with the highest score in the overlapping area of ​​the image slices and the traversed candidate detection box respectively.

[0248] Step S7.3: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than a threshold, the currently traversed detection frame is deleted;

[0249] The above process is repeated to retain the valid detection frames in the overlapping areas, thereby eliminating the problem of multiple detections of the same spatial target in adjacent image slices.

[0250] Table 2 shows the detection frames in the original image mapped to adjacent detection image slices. The detection frames traversed in steps 7.1 to 7.4 are compared with the frame with the highest score to calculate the intersection and union ratio. The results are shown in Table 3. For a pair of detection frames with an intersection and union ratio greater than 0.5, the detection frame with low confidence is deleted and the detection frame with high confidence is retained to achieve the screening and merging of detection results.

[0251] Table 2. Detection boxes in the original image mapped to adjacent sub-images;

[0252]

[0253] Table 3. Intersection over Union (IoU) between adjacent sub-image detection boxes

[0254] IoU Detection frame 1-1 Detection frame 1-2 Detection frame 1-3 Detection frame 2-1 0.72 0 0 Detection frame 2-2 0.09 0.57 0 Detection frame 2-3 0 0 0.64

[0255] The present invention realizes star background suppression and space target detection in space detection images, is applicable to multi-frame observation images under different detection modes, and improves the accuracy and universality of the detection method.

[0256] The present invention also provides a spatial detection image dim target detection system that combines time and space domains. The spatial detection image dim target detection system that combines time and space domains can be realized by executing the process steps of the spatial detection image dim target detection method that combines time and space domains. That is, those skilled in the art can understand the spatial detection image dim target detection method that combines time and space domains as a preferred implementation of the spatial detection image dim target detection system that combines time and space domains.

[0257] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0258] According to the present invention, a spatial detection image dim target detection system combined with time and space domains includes:

[0259] Module M1: Slice the original space target detection image with equal steps and overlapping areas to obtain the image slices to be detected after cutting;

[0260] Module M2: Extract key points for each image slice;

[0261] Module M3: construct the topological space description features of key points, and perform inter-frame registration of adjacent frame images based on the key point extraction results of adjacent frames;

[0262] Module M4: Calculate the inter-frame difference of adjacent frames after registration to obtain the inter-frame difference result;

[0263] Module M5: superimpose the inter-frame difference results to obtain the background suppression result;

[0264] Module M6: Perform spatial target training and detection on the background suppression results based on the Yolov5 deep learning network to obtain the detection results;

[0265] Module M7: Filter and merge the test results and output the final test results.

[0266] In more preferred embodiments, the module M2 includes:

[0267] Module M2.1: Binarize the image slices to obtain a binary image;

[0268] Module M2.2: Calculate the connected domain of stars in the binary image, where the connected domain is an eight-neighborhood domain;

[0269] Module M2.3: Find the centroid of the connected domain by grayscale weighted calculation, according to The obtained centroid is taken as the key point;

[0270] Among them, y′ represents the coordinates of the centroid row vector;

[0271] z′ represents the centroid column vector coordinates;

[0272] y″ represents the row coordinate of the i-th pixel in the connected domain;

[0273] z″ represents the column coordinates of the i-th pixel in the connected domain;

[0274] h represents the gray value of the i-th pixel in the connected domain;

[0275] N represents a total of N pixels in the connected domain.

[0276] The module M3 comprises:

[0277] Module M3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points;

[0278] Module M3.2: Perform key point matching to obtain matching point pairs;

[0279] The matching is to calculate the feature similarity of key points in the current frame and the adjacent frames;

[0280] Module M3.3: align adjacent frames to the starry sky background based on the current frame according to calculation;

[0281] The calculation is to obtain the transformation matrix from the adjacent frame image to the current frame image based on the matched point pairs based on homography transformation and estimation;

[0282] The homography transformation and estimation formula are:

[0283] in, represents the homography matrix;

[0284] h ij represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3;

[0285] x represents the horizontal coordinate of the key point in the adjacent frame image;

[0286] y represents the vertical coordinate of the key point in the adjacent frame image;

[0287] x′ is the horizontal coordinate of the same key point in the current frame;

[0288] y′ represents the ordinate of the corresponding key point in the current frame;

[0289] The module M6 comprises:

[0290] Module M6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters;

[0291] Module M6.2: Use the trained network parameters to detect the space target detection image and obtain the candidate detection box;

[0292] The Yolov5 deep learning network is a YOLOv5s model;

[0293] The module M7 comprises:

[0294] Module M7.1: Sort the obtained multiple candidate detection boxes according to the confidence scores;

[0295] Module M7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them. The formula is

[0296] Among them, IoU represents the degree of overlap;

[0297] A and B represent the candidate detection box with the highest score and the traversed candidate detection box in the overlapping area of ​​the image slices, respectively;

[0298] Module M7.3: Judgment: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than the threshold, the currently traversed candidate detection frame is deleted; when the intersection-and-union ratio is less than the threshold, the candidate detection frame is retained;

[0299] Modules M7.1 to M7.3 are triggered repeatedly until all candidate detection frames are traversed and the process stops.

[0300] In more preferred embodiments, the module M3.1 includes:

[0301] Module M3.1.1: Take any key point p i As the center of the circle, a set number of concentric circles are constructed around it and divided equally into a set number of regions;

[0302] Module M3.1.2: For key point p i The key points around j , p j The coordinates in the Cartesian coordinate system are transformed into the key point pi is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j , the corresponding area number q j for:

[0303] 0≤θ j ≤2π,0≤ρ j ≤R;

[0304] Where R represents the maximum concentric circle radius;

[0305] m represents the unit radian when quantizing the polar angle;

[0306] n is the unit distance when quantizing the polar diameter;

[0307] Module M3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature F i ;

[0308] The module M3.2 includes:

[0309] Module M3.2.1: Set the current frame I n The key point p i The position in the Cartesian coordinate system (x i ,y i ) as the search starting point, in the adjacent frame I n +Δ searches for a key point p near the starting point k And calculate its position feature F k ;

[0310] Module M3.2.2: Calculate the distance between position features and the grayscale difference of key points in the original image slice;

[0311] Module M3.2.3: If the formula conditions are met, the current frame I n The key point p i and adjacent frame I n +Δ key point p k If the match is successful, it will be regarded as a pair of matched points; if it is not satisfied, the calculation will continue;

[0312] The formula conditions are

[0313] Among them, q represents the region number in the feature;

[0314] f(p i ) represents the key point p i Grayscale value on the original image;

[0315] f(p k ) represents the key point p k Grayscale value on the original image;

[0316] th1 is the threshold of the distance between features;

[0317] th2 is the grayscale difference threshold;

[0318] Traverse all key points and trigger modules M3.2.1 to M3.2.3 to obtain all matching point pairs.

[0319] In more preferred embodiments, the formula calculated in the module M4 is:

[0320]

[0321] Among them, Sat_Value represents the saturation value of the image;

[0322] D N (i, j) represents the difference result between adjacent frames at position (i, j);

[0323] f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame;

[0324] f N-2 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame is registered.

[0325] In the module M5, the current frame is registered and differentiated with multiple adjacent frames, and the registration difference result obtained after the registration difference from the jth frame to the kth frame is D(I j ,I j-k ), by accumulating n frames between frames, taking the average value of n frame registration difference results, the formula is

[0326] Among them, I j represents the spatial target signal in the jth frame;

[0327] I j-1 Represents the spatial target signal from the jth frame to the kth frame;

[0328] I j ' represents the enhanced spatial target signal in the jth frame.

[0329] According to a computer-readable storage medium storing a computer program provided by the present invention, when the computer program is executed by a processor, the steps of the method for detecting dim targets in spatial detection images in a combined time-space domain are implemented.

[0330] An electronic device provided according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the method for detecting dim targets in spatial detection images by combining time and space domains are implemented.

[0331] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for detecting dim targets in spatial detection images in a combined temporal and spatial domain, characterized in that: include: Step S1: Slice the original space target detection image with equal step length and overlapping areas to obtain the image slices to be detected after cutting; Step S2: extract key points for each image slice; Step S3: constructing the topological space description features of the key points, and performing inter-frame registration of adjacent frame images according to the key point extraction results of adjacent frames; Step S4: Calculate the inter-frame difference of adjacent frames after registration to obtain the inter-frame difference result; Step S5: superimposing the inter-frame difference results to obtain a background suppression result; Step S6: Perform spatial target training and detection on the background suppression result through the Yolov5 deep learning network to obtain the detection result; Step S7: Screen and merge the test results and output the final test results.

2. The method for detecting dim targets in spatial detection images by combining time and space domains according to claim 1, characterized in that: The step S2 comprises: Step S2.1: binarizing the image slice to obtain a binary image; Step S2.2: Calculate the connected domain of the stars in the binary image, where the connected domain is an eight-neighborhood domain; Step S2.3: Calculate the centroid of the connected domain by grayscale weighted calculation, according to The obtained centroid is taken as the key point; Among them, y′ represents the coordinates of the centroid row vector; z′ represents the centroid column vector coordinates; y″ represents the row coordinate of the i-th pixel in the connected domain; z″ represents the column coordinates of the i-th pixel in the connected domain; h represents the gray value of the i-th pixel in the connected domain; N represents a total of N pixels in the connected domain; The step S3 comprises: Step S3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points; Step S3.2: performing key point matching to obtain matching point pairs; The matching is to calculate the feature similarity of key points in the current frame and the adjacent frames; Step S3.3: aligning adjacent frames to the starry sky background based on the current frame according to calculation; The calculation is to obtain the transformation matrix from the adjacent frame image to the current frame image based on the matched point pairs based on homography transformation and estimation; The homography transformation and estimation formula are: in, represents the homography matrix; h &' represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3; x represents the horizontal coordinate of the key point in the adjacent frame image; y represents the vertical coordinate of the key point in the adjacent frame image; x′ is the horizontal coordinate of the same key point in the current frame; y′ represents the ordinate of the corresponding key point in the current frame; The step S6 comprises: Step S6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters; Step S6.2: Detect the space target detection image using the trained network parameters to obtain a candidate detection frame; The Yolov5 deep learning network is a YOLOv5s model; The step S7 comprises: Step S7.1: sorting the obtained multiple candidate detection boxes according to the confidence scores; Step S7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them. The formula is: Among them, IoU represents the degree of overlap; A and B represent the candidate detection box with the highest score and the traversed candidate detection box in the overlapping area of ​​the image slices, respectively; Step S7.3: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than the threshold, the currently traversed candidate detection frame is deleted; when the intersection-and-union ratio is less than the threshold, the candidate detection frame is retained; Repeat steps S7.1 to S7.3 until all candidate detection frames have been traversed.

3. The method for detecting dim targets in spatial detection images in the spatiotemporal domain combined with the temporal domain according to claim 2 is characterized in that: The step S3.1 comprises: Step S3.1.1: Take any key point p i As the center of the circle, a set number of concentric circles are constructed around it and divided equally into a set number of regions; Step S3.1.2: For key point p i The key points around j , p j The coordinates in the Cartesian coordinate system are transformed into the key point p i is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j , the corresponding area number q j for: Where R represents the maximum concentric circle radius; m represents the unit radian when quantizing the polar angle; n is the unit distance when quantizing the polar diameter; Step S3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature F i ; The step S3.2 comprises: Step S3.2.1: Set the current frame I n The key point p i The position in the Cartesian coordinate system (x & ,y & ) as the search starting point, in the adjacent frame I n +Δ searches for a key point p near the starting point k And calculate its position feature F k ; Step S3.2.2: Calculate the distance between the position features and the grayscale difference of the key points in the original image slice; Step S3.2.3: If the formula condition is met, the current frame I n The key point p i and adjacent frame I n +Δ key point p k If the match is successful, it will be regarded as a pair of matched points; if it is not satisfied, the calculation will continue; The formula conditions are Where q represents the region number in the feature; f+p i , represents the key point p i Grayscale value on the original image; f+p k , represents the key point p k Grayscale value on the original image; th1 is the threshold of the distance between features; th2 is the threshold of grayscale difference; Traverse all key points and execute steps S3.2.1 to S3.2.3 to obtain all matching point pairs.

4. The method for detecting dim targets in spatial detection images by combining time and space domains according to claim 1, characterized in that: The formula calculated in step S4 is: Among them, Sat_Value represents the saturation value of the image; D N (i, j) represents the difference result between adjacent frames at position (i, j); f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame; f N-1 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame registration. In step S5, the current frame is registered and differentiated with multiple adjacent frames, and the registration difference result obtained after the registration difference between the jth frame and the kth frame is D(I j , I j-k ), by accumulating n frames between frames, taking the average value of n frame registration difference results, the formula is Among them, I j represents the spatial target signal in the jth frame; I j-k Represents the spatial target signal from the jth frame to the kth frame; I j ” represents the enhanced spatial target signal in the jth frame.

5. A spatial detection image dim target detection system combined with time and space domains, characterized in that: include: Module M1: Slice the original space target detection image with equal steps and overlapping areas to obtain the image slices to be detected after cutting; Module M2: Extract key points for each image slice; Module M3: construct the topological space description features of key points, and perform inter-frame registration of adjacent frame images based on the key point extraction results of adjacent frames; Module M4: Calculate the inter-frame difference of adjacent frames after registration to obtain the inter-frame difference result; Module M5: superimpose the inter-frame difference results to obtain the background suppression result; Module M6: Based on the Yolov5 deep learning network, the background suppression results are trained and detected to obtain the detection results; Module M7: Filter and merge the test results and output the final test results.

6. The spatial detection image dim target detection system of the time-space domain joint spatial detection according to claim 5 is characterized in that: The module M2 comprises: Module M2.1: Binarize the image slices to obtain a binary image; Module M2.2: Calculate the connected domain of stars in the binary image, where the connected domain is an eight-neighborhood domain; Module M2.3: Find the centroid of the connected domain by grayscale weighted calculation, according to The obtained centroid is taken as the key point; Among them, y′ represents the coordinates of the centroid row vector; z′ represents the centroid column vector coordinates; y″ represents the row coordinate of the i-th pixel in the connected domain; z′′ represents the column coordinates of the i-th pixel in the connected domain; h represents the gray value of the i-th pixel in the connected domain; N represents a total of N pixels in the connected domain. The module M3 comprises: Module M3.1: Taking each key point as the center, construct the topological space description features of it and the surrounding key points; Module M3.2: Perform key point matching to obtain matching point pairs; The matching is to calculate the feature similarity of key points in the current frame and the adjacent frames; Module M3.3: align adjacent frames to the starry sky background based on the current frame according to calculation; The calculation is to obtain the transformation matrix from the adjacent frame image to the current frame image based on the matched point pairs based on homography transformation and estimation; The homography transformation and estimation formula are: in, represents the homography matrix; h ij represents the corresponding elements in the homography matrix, i = 1, 2, 3, j = 1, 2, 3; x represents the horizontal coordinate of the key point in the adjacent frame image; y represents the vertical coordinate of the key point in the adjacent frame image; x′ is the horizontal coordinate of the same key point in the current frame; y′ represents the ordinate of the corresponding key point in the current frame; The module M6 comprises: Module M6.1: Generate a large number of space target detection images through simulation, build a training data set, and use space target labels to tune and train the Yolov5 deep learning network to obtain network parameters; Module M6.2: Use the trained network parameters to detect the space target detection image and obtain the candidate detection box; The Yolov5 deep learning network is a YOLOv5s model; The module M7 comprises: Module M7.1: Sort the obtained multiple candidate detection boxes according to the confidence scores; Module M7.2: For the candidate detection frame with the highest score, traverse the remaining candidate detection frames in turn and calculate the intersection-over-union ratio between them. The formula is Among them, IoU represents the degree of overlap; A and B represent the candidate detection box with the highest score and the traversed candidate detection box in the overlapping area of ​​the image slices, respectively; Module M7.3: Judgment: When the intersection-and-union ratio of the traversed candidate detection frame and the candidate detection frame with the highest score is greater than the threshold, the currently traversed candidate detection frame is deleted; when the intersection-and-union ratio is less than the threshold, the candidate detection frame is retained; Modules M7.1 to M7.3 are triggered repeatedly until all candidate detection frames are traversed and the process stops.

7. The spatial detection image dim target detection system of the time-space domain joint spatial detection according to claim 6 is characterized in that: The module M3.1 includes: Module M3.1.1: Take any key point p i As the center of the circle, a set number of concentric circles are constructed around it and divided equally into a set number of regions; Module M3.1.2: For key point p i The key points around j , p j The coordinates in the Cartesian coordinate system are transformed into the key point p i is the polar coordinate of the origin, and its polar diameter ρ is obtained j and the polar angle θ j , the corresponding area number q j for: Where R represents the maximum concentric circle radius; m represents the unit radian when quantizing the polar angle; n is the unit distance when quantizing the polar diameter; Module M3.1.3: Count the number of key points falling into each area around the key point, and sort them according to the area number to obtain the dimensional vector corresponding to the key point As the key point p i The position feature F i ; The module M3.2 includes: Module M3.2.1: Set the current frame I n The key point p i The position in the Cartesian coordinate system (x & ,y & ) as the search starting point, in the adjacent frame I n +Δ searches for a key point p near the starting point k And calculate its position feature F k ; Module M3.2.2: Calculate the distance between position features and the grayscale difference of key points in the original image slice; Module M3.2.3: If the formula conditions are met, the current frame I n The key point p i and adjacent frame I n +Δ key point p k If the match is successful, it will be regarded as a pair of matched points; if it is not satisfied, the calculation will continue; The formula conditions are Among them, q represents the region number in the feature; f(p i ) represents the key point p i Grayscale value on the original image; f(p k ) represents the key point p k Grayscale value on the original image; th1 is the threshold of the distance between features; th2 is the threshold of grayscale difference; Traverse all key points and trigger modules M3.2.1 to M3.2.3 to obtain all matching point pairs.

8. The spatial detection image dim target detection system of the time-space domain joint according to claim 5 is characterized in that: The calculation formula in the module M4 is: Among them, Sat_Value represents the saturation value of the image; D N (i, j) represents the difference result between adjacent frames at position (i, j); f N (i, j) represents the grayscale value at position (i, j) of the current Nth frame; f N-1 (i, j) represents the gray value at the corresponding position (i, j) after the previous frame registration; In the module M5, the current frame is registered and differentiated with multiple adjacent frames, and the registration difference result obtained after the registration difference from the jth frame to the kth frame is D(I j , I j-k ), by accumulating n frames between frames, taking the average value of n frame registration difference results, the formula is Among them, I j represents the spatial target signal in the jth frame; I j-k Represents the spatial target signal from the jth frame to the kth frame; I j ' represents the enhanced spatial target signal in the jth frame.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for detecting dim targets in spatial detection images by combining time and space domains as described in any one of claims 1 to 4 are implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the method for detecting dim targets in spatial detection images by combining time and space domains as described in any one of claims 1 to 4 are implemented.

Images