Data-driven point-like space object detection method

Abstract

The application discloses a point-shaped space target detection method based on data driving, which comprises the following steps: pre-processing a to-be-detected image to obtain a pre-processed image; searching and processing the pre-processed image according to the minimum pixel level radius of the image to obtain a candidate region; and inputting the candidate region to an SF-CNN classification model as data driving to guide classification, so as to obtain an image prediction category. Through the application, the detection precision of the image can be improved while reducing the workload of global labeling by detecting each candidate region of the image. The application can be widely applied to the technical field of deep learning target detection.

Description

Technical Field

[0001] This invention relates to the field of deep learning object detection technology, and in particular to a data-driven point spatial object detection method. Background Technology

[0002] During celestial observations, rapid and accurate detection of space targets in optical images is a crucial issue that must be addressed. Due to the vast imaging distance and short exposure times, space targets and stellar targets appear as point-like distributions in images. Detecting such targets presents several challenges: First, the targets are small and lack geometric, structural, and textural features. Second, during observation, space targets may flip, and observation conditions such as the distance between the target and the optical system, and the angle of incidence of the sun relative to the target, are dynamically changing, altering the grayscale values ​​of the space targets in the sequence of star images. Third, to better reflect the grayscale differences between targets and the grayscale changes of individual targets, images with higher depth values ​​are often used in engineering, resulting in significant differences between images and targets. Finally, when the optical system is in orbit, limited hardware resources place high demands on the computational complexity of the algorithm. Most existing methods rely on background suppression and target saliency enhancement for target detection, neglecting the target's inherent distribution characteristics. This leads to a lack of robustness when dealing with targets exhibiting large variations in grayscale and size, or strong interference. Furthermore, current model-driven methods have limitations when targets deviate from their ideal distribution. This is because high relative speeds between the target and the platform, or long exposure times, cause motion blur, extending the target's distribution along the motion trajectory. Since different target motion directions and speeds result in variations in the direction and length of this extension in the image, making it difficult to generalize a single objective function for different scenarios, even with parameter normalization during target detection. While methods such as image processing exist, the threshold setting issue remains. Currently, more and more scholars are applying deep learning as a typical data-driven approach to object detection. However, existing deep learning methods are difficult to directly apply to spatial object detection due to the following problems: First, even when photographing the same sky area, cameras with different parameters have different detection capabilities, meaning the number and location of stars and targets in the captured images cannot be determined, making it extremely difficult to construct a global dataset. Second, commonly used network structures contain a large number of pooling layers, making them difficult to directly apply to the detection of weak targets. Third, deep learning algorithms typically require a large amount of computational resources, which places extremely high demands on the system's hardware. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a data-driven point spatial target detection method that can improve image detection accuracy by detecting each candidate region of an image while reducing the workload of global annotation.

[0004] The first technical solution adopted in this invention is: a data-driven point spatial target detection method, comprising the following steps:

[0005] The image to be detected is preprocessed to obtain the preprocessed image;

[0006] The preprocessed image is searched based on the smallest pixel-level radius to obtain candidate regions.

[0007] The candidate regions are used as data-driven inputs to the SF-CNN classification model for guided classification, resulting in the predicted image category.

[0008] Furthermore, the step of preprocessing the image to be detected to obtain a preprocessed image specifically includes:

[0009] The star map is photographed using a monitoring telescope to obtain the image to be detected;

[0010] The image to be detected is processed by a median filter to remove noise, resulting in a noise-suppressed image.

[0011] The noise-suppressed image is cropped and the background value of the image to be detected is estimated to obtain the block background value;

[0012] The background values ​​of the blocks are transformed using bilinear interpolation to obtain the background image and the root mean square image.

[0013] The background image is segmented to obtain the foreground image, which includes the star to be detected and the target.

[0014] The foreground image is binarized and encoded sequentially based on the root mean square plot to obtain the preprocessed image.

[0015] Furthermore, the step of cropping the noise-suppressed image and estimating the background value of the image to be detected to obtain the block background value specifically includes:

[0016] The image after noise suppression is iteratively cropped until the background value of the image meets the preset convergence range, at which point the cropping is stopped and the cropped image is obtained.

[0017] Determine changes in the image background value during the cropping process;

[0018] If it is determined that the rate at which the background value of the image decreases in each cropping is less than the preset rate of decrease, the average value of the background value of the cropped image is selected as the block background value.

[0019] If it is determined that the rate of decrease of the background value of the cropped image is greater than the preset rate of decrease, the pattern estimation is performed on the background value of the cropped image to obtain the block background value.

[0020] Furthermore, the calculation process for the mode estimation is as follows:

[0021] mode = 2.5 × median - 1.5 × mean

[0022] In the above formula, mode represents mode estimation, median represents the output median of the median filter, and mean represents the mean.

[0023] Furthermore, the step of searching the preprocessed image based on the smallest pixel-level radius to obtain candidate regions specifically includes:

[0024] The minimum search distance between targets in the preprocessed image is determined based on the minimum pixel-level radius of the image.

[0025] The dilation operation region is determined based on the minimum search distance between targets in the preprocessed image;

[0026] The preprocessed image is subjected to dilation search processing based on the dilation operation region to obtain the local maximum point of the image;

[0027] Candidate regions are obtained by constructing a set of regions containing local maximum points in the image.

[0028] Furthermore, the step of using candidate regions as data-driven inputs to the SF-CNN classification model for guided classification to obtain the image prediction category specifically includes:

[0029] The candidate regions are normalized to obtain normalized candidate regions;

[0030] Image interpolation is performed on the normalized candidate regions to obtain the interpolated candidate regions.

[0031] The interpolated candidate regions are input into the SF-CNN classification model for classification to obtain classification results, which include confidence values ​​for targets and confidence values ​​for non-targets.

[0032] The size of the target is determined based on the classification results, and corresponding saliency and radius maps are generated.

[0033] Redundant targets are eliminated from the saliency map and radius map using a non-maximum suppression method to obtain the image prediction category.

[0034] Furthermore, the expression for the nonmaximum suppression method is as follows:

[0035]

[0036]

[0037] In the above formula, V R The value of V within radius R represents the maximum value of V, where V represents the significance plot and R represents the radius plot. i ,y i V(x) represents the suspected center point. i ,y i ) represents a point (x) on the saliency plot. i ,y i ), R(x i ,y i ) represents a point (x) on the radius graph. i ,y i ).

[0038] The step of inputting the interpolated candidate region into the SF-CNN classification model for classification to obtain the classification result specifically includes:

[0039] The interpolated candidate regions are input into the SF-CNN classification model, which includes a first convolutional layer, a pooling layer, a second convolutional layer, and a fully connected layer.

[0040] Based on the first convolutional layer, feature extraction processing is performed on the interpolated candidate regions to obtain the image's feature information;

[0041] The pooling layer is used to downsample the image feature information to obtain low spatial dimension image feature information.

[0042] The image feature information with low spatial dimension is combined and processed based on the second convolutional layer to obtain the combined image feature information.

[0043] Based on the fully connected layer, the confidence value of the interpolated candidate region is calculated and classified according to the combined image feature information, and the classification result is output.

[0044] Furthermore, the loss function of the second convolutional layer is the bias loss function, the specific expression of which is shown below:

[0045]

[0046]

[0047] In the above formula, L bias Represents the deviation loss function. The function represents a nonlinear scaling function, where α and β represent adjustable contribution parameters, and q... ij Indicates whether the i-th sample belongs to category j, p i Let f represent the feature value corresponding to the i-th sample, θ represent the model parameters, and f j (·) represents a neural network. This represents the variance of the i-th sample after processing by the convolutional layer output, and N represents the total number of samples.

[0048] The beneficial effects of the method of this invention are as follows: This invention acquires star maps through observation and photography using a monitoring telescope, further preprocesses the star maps to eliminate thermal noise and smear effects, incorporates a pattern estimation method to eliminate image noise during the noise elimination process, and can effectively suppress image noise under excessive interference. Furthermore, it applies the imaging features of targets and stars to search for target candidate regions in the image, and uses the processed regions as input to the SF-CNN model. It adopts a data-driven approach to replace the traditional model-driven approach, making the algorithm applicable to more situations. The SF-CNN model is a simplified classification model that transforms the detection problem across the entire image into a local classification problem, and can identify each candidate region to achieve the full-image target detection problem. Attached Figure Description

[0049] Figure 1 This is a flowchart of the steps of the data-driven point spatial target detection method of the present invention;

[0050] Figure 2 This is a schematic diagram showing the data comparison obtained by pattern estimation of the background of the cropped image according to the present invention;

[0051] Figure 3 This is a schematic diagram of the method for detecting and identifying candidate regions based on the SF-CNN model of the present invention;

[0052] Figure 4 This is a schematic diagram of the process for extracting the center portion of the candidate region according to the present invention;

[0053] Figure 5 This is a schematic diagram illustrating the steps of fully connecting the central portion of the extracted candidate region in this invention;

[0054] Figure 6 This is a schematic diagram of the normalized negative sample star map obtained by the present invention;

[0055] Figure 7 This is a schematic diagram of the normalized positive sample star map obtained by the present invention;

[0056] Figure 8 This is a line graph of the loss curve during model training in this invention;

[0057] Figure 9 This is a schematic diagram of the original sub-images and corresponding grayscale distributions of the same target under different saliency conditions according to the present invention;

[0058] Figure 10 This is a comparative schematic diagram of the image before and after preprocessing according to the present invention;

[0059] Figure 11 This is a comparative schematic diagram of the preprocessing of stellar images before and after the present invention;

[0060] Figure 12 This is a saliency map and a schematic diagram of the detection results of image detection using existing model algorithms and the SF-CNN model of this invention;

[0061] Figure 13 This is a schematic diagram illustrating the target missed detection results using existing model algorithms and the SF-CNN model of this invention.

[0062] Figure 14 This is a schematic diagram comparing the ROC curves of existing model algorithms and the SF-CNN model of this invention under different sequences;

[0063] Figure 15 These are saliency maps and detection results obtained under different saliency levels based on the same target using existing model algorithms and the SF-CNN model of this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0065] Reference Figure 1 This invention provides a data-driven point-based spatial target detection method, which includes the following steps:

[0066] S1. Preprocess the image to be detected to obtain the preprocessed image;

[0067] Specifically, during space monitoring, there may be some interference, such as thermal noise and smear effects. Therefore, a median filter is used to eliminate them while preserving the distribution characteristics of the target. The calculation formula is as follows:

[0068] f n (x,y)=med(src(x+x′,y+y′)|(x′,y′)∈D med )

[0069] In the above formula, src(x, y) represents the original image, and f m (x, y) represents the filtered image, D med Indicates the median filtering region;

[0070] Furthermore, considering the differences in PSF effects caused by the optical system, the D values ​​from 3×3 to 7×7... med A 3×3 area typically yields good results; in this invention, a 3×3 area is chosen.

[0071] Considering the potential for excessive interference, this chapter introduces a method based on SExtractor. This method has a good effect on suppressing background noise and is suitable for the research environment in this invention. It utilizes a mesh to divide f... m The block is divided into several blocks. For each block, the background is estimated by combining k.σ-clipping and mode estimation. First, the block background histogram is iteratively clipped until the background value converges within ±3σ of the median. Second, if σ decreases at a rate of less than 20% during the iteration, the average of the clipped histogram is directly selected as the block background value. Otherwise, the background of this block is considered "crowded" and a new mode that better conforms to the clipping distribution is used to estimate the block background. The calculation formula is as follows:

[0072] mode = 2.5 × median - 1.5 × mean

[0073] In the above formula, mode represents mode estimation, median represents the output median of the median filter, and mean represents the mean.

[0074] like Figure 2 As shown, assuming the block size is 32×32 and is contaminated by a Gaussian distribution, its true value is close to 0. The points in the image are the calculation results of the two methods after each cropping. It can be seen that the above mode gives a more robust estimate than the simple mean and is much less affected by the "crowding" effect.

[0075] The calculated block background is transformed into a background image f using bilinear interpolation. bkg Root Mean Square (RMS) plot f RMS Typically, a grid size between 32 and 128 is selected; here, it is set to 32. The foreground f... fkg This mainly includes stars and targets, where targets refer to space objects such as space debris, satellites, etc., and are identified from the background f in the following manner. bkg It is separated from the middle, and its expression is:

[0076] f fkg (x, y) = f(x, y) - f bkg (x, y)

[0077] In the above formula, f fkg (x, y) represents the foreground image, f bkg (x, y) represents the background image;

[0078] The RMS image is directly applied to the binarization of the foreground, and its calculation process is as follows:

[0079]

[0080] In the above formula, f bw (x, y) represents the binarized image. For the binarized image, the two-pass algorithm is used to encode each connected region and use it as a potential target region.

[0081] S2. Search the preprocessed image based on the minimum pixel-level radius of the image to obtain candidate regions;

[0082] Specifically, the minimum pixel-level radius of the image is obtained based on statistical data and empirical values. Based on the correlation between the target center and local maxima in the image, a dilation operation is used to quickly extract the local maxima in the image. The extraction process is as follows:

[0083] f p (x, y) = max(f m (x+x′,y+y′)|(x′,y′)∈D m )

[0084] In the above formula, f p (x, y) represents the extended image, D m This represents the expansion operation area determined based on the minimum search distance between targets;

[0085] Since it is difficult to distinguish between a target with a radius of 1 and thermal noise, the smallest pixel-level radius at which one can determine whether a region contains a target is 2. Therefore, this invention will use D m Set to 5 (=2×2+1) to maintain the method's sensitivity to two close targets, the suspected center point. The point set is defined as follows:

[0086]

[0087] In the above formula, for each point Regions of different sizes in the surrounding area will be extracted as candidate regions. Smaller Dm may result in more candidate regions that only contain the background, so these candidate regions will be classified in the next subsection.

[0088] S3. The candidate region is used as data-driven input to the SF-CNN classification model for guided classification to obtain the image prediction category.

[0089] Specifically, refer to Figure 3 Due to the limitations of hardware resources and computing power in orbit, this invention designs a simple and small SF-CNN classification model to identify each candidate region and thus achieve the problem of full image target detection, instead of directly using the network structure in existing deep learning methods. The SF-CNN classification model is trained to obtain a small model C, which classifies candidate regions into targets or non-targets.

[0090] The model's input is a processed candidate region f norm The output is the predicted category (i.e., target (=1) or non-target (=0)) and the corresponding confidence value v. c (The output value of C), when the loss is concentrated at 0.25, training of the parameters will stop, at which point the trained C is considered to have learned the distribution of small targets. When the distribution of targets is not severely affected by the background or stars, C can be directly used to predict the input category;

[0091] Since the size of the target cannot be determined in advance, for each suspected center point (x i y i Candidate regions of different sizes k are extracted to confirm the size of the target. The saliency map V and radius map R are generated as follows:

[0092]

[0093] V(x i y i ) = max(V c (x i y i ))

[0094] R(x i y i )={k|V k (x i y i )=V(x i y i )}

[0095] k = 5, 6, ..., k end -1,k end

[0096] In the above formula, V c (x i y i (x, k) represents the point (x, k) with size k. i y i ) corresponding confidence level It should be noted that if the center of the candidate region (the suspected target center) is not the maximum value within the region at size k, then... Set to 0, σ end The setting needs to be combined with the information of the optical system, and this invention selected 11;

[0097] The saliency map is used to determine the confidence level of the target, i.e., to compare which suspected point is more likely to be the target; the radius map is used to select the target, and its radius is obtained based on statistical results and empirical values.

[0098] To eliminate redundant targets in the same candidate region, non-maximum suppression (NMS) is employed, and its expression is shown below:

[0099]

[0100]

[0101] In the above formula, V R The value of V within radius R represents the maximum value of V, where V represents the significance plot and R represents the radius plot. i y i V(x) represents the suspected center point. i y i ) represents a point (x) on the saliency plot. i y i ), R(x i y i ) represents a point (x) on the radius graph. i y i );

[0102] S31, Network Input;

[0103] Specifically, in space surveillance and astronomical observation, 16-bit optical images are typically chosen because they have a wider grayscale range than 8-bit images, better representing the details within the image. To minimize the impact of grayscale differences and lighting conditions while preserving the target distribution as much as possible, each candidate region undergoes normalization processing, as shown below:

[0104]

[0105]

[0106] In the above formula, f sub (x, y) represents the original candidate region, (x...y) i y i ) represents the coordinates of the i-th suspected center point, k represents the size of the candidate region, and f norm(x, y) represents the normalized candidate region;

[0107] Since the input image size of the network is constant, regions of different scales will be scaled to the same size. When performing normalization on the same target at different scales, if the target center is the maximum value in the region, the normalization has little impact on the target distribution. Therefore, in order to fully extract information from the input image, a sufficiently large size needs to be selected to facilitate multiple subsequent convolutions of the image, thereby achieving the purpose of feature extraction. At the same time, it is also necessary to consider that an excessively large size may cause other targets or interference in the non-target region of the image to have too much influence. Considering the above factors, the image interpolation method is chosen to scale regions of different scales to a size of 11×11 as the input of the network.

[0108] S32, Network Structure;

[0109] Specifically, the detailed structure of the SF-CNN network designed in this invention is shown in Table 1;

[0110] Table 1. SF-CNN Network Structure

[0111]

[0112] The process involves using convolutional layers to extract multi-level features from the image. Neurons in the first convolutional layer are highly sensitive to very local image structures such as edges, while neurons in deeper convolutional layers can combine lower-level features more effectively. Max pooling layers downsample the feature maps while keeping the features unchanged, thereby reducing the spatial dimension and creating a larger receptive field in subsequent convolutional layers. Fully connected layers identify the category of the input image and output the corresponding confidence score.

[0113] Since the target is located in the center of the input image, to further reduce interference from pixels in non-central areas and enhance the target's distribution characteristics, a 5×5 region at the center of the candidate region is extracted and input into the Feature Extractor (FE) module. This extractor utilizes three feature parameters E... R E C and E T The three feature parameters are energy responsiveness, energy concentration, and energy transfer. The calculation results are used as guidance information and added to the output of Convolution-2 (the second convolution) for subsequent full connection.

[0114] The network's input is a processed sub-image. This network can learn the distribution characteristics of the target from the input data and adjust its parameters to build a target model. The network's output consists of two neurons: one representing the confidence value of the target, and the other representing the confidence value of the non-target. Figure 4 and Figure 5 The document details the model and the output of the intermediate processes. Figure 4 The numbers in the diagram represent the specifications of the input image. During training, the weights and biases of these parameters are learned through backpropagation. After training, the network's loss is around 0.25, at which point the distribution characteristics of the targets in the sample data are considered to have been sufficiently learned by the network.

[0115] S33. Calculation of the loss function;

[0116] Specifically, considering that data points are unlikely to provide a sufficient number of unique features to well describe the target, this problem can lead to misleading information in the optimization process caused by random prediction. This invention introduces bias loss to solve resource-constrained classification problems, that is, to focus training on valuable data points and prevent a large number of poorly learned samples from misleading the optimization process. In this process, variance can be used as a simple measure of diversity to reflect the distance between the value of the feature map and the mean. The larger the variance, the greater the chance of obtaining a large number of unique features. The feature map of the last convolutional layer will be used for variance calculation.

[0117] Assumption It is the output of the convolutional layer, which unfolds T into a two-dimensional array. Batch refers to the number of input data points required to calculate a single loss. The formula for calculating the variance of the feature map of the i-th data point is as follows:

[0118]

[0119]

[0120] n = c × h × w

[0121] In the above formula, t j Let represent the j-th output, b represent the batch size, c represent the number of input channels, and h and w represent the height and width of the tensor.

[0122] In addition, variance is processed as follows:

[0123]

[0124] In each iteration, v max and v minThese are the maximum and minimum activation values ​​in a set of feature maps. This is done to ensure that outliers in the variance do not cause significant changes in the loss, thereby maintaining model stability;

[0125] set up Let Q be the feature space, and Q = {0, 1} be the class space. To calibrate the contribution of each data point to the cumulative loss, for the dataset... (where p) i With q i These are the feature value and category corresponding to the i-th data point, respectively. i ,q i )∈P×Q represents) and the neural network f(x; θ) (where θ represents the model parameters). Typically, the cross-entropy loss, commonly used in classification networks, is defined as:

[0126]

[0127] A non-linear scaling function z is added to the traditional cross-entropy loss to introduce bias between data points with low variance and those with high variance. The bias loss is defined as follows:

[0128]

[0129]

[0130] In the above formula, L bias Represents the deviation loss function. The nonlinear scaling function is represented by α and β, which are adjustable contribution parameters. In this invention, α = 0.5 and β = 0.3 are selected. ij Indicates whether the i-th sample belongs to category j, p i Let f represent the feature value corresponding to the i-th sample, θ represent the model parameters, and f j (·) represents a neural network. This represents the variance of the i-th sample after processing by the convolutional layer output, and N represents the total number of samples.

[0131] The simulation experiments of this invention are as follows:

[0132] The above method was applied to spatial object detection. The CNN was implemented in Python 3.7 on a PC with 16GB of memory, a 1.9GHz Intel i7 CPU and an NVIDIA GeForce GTX 1060 GPU. During the experiment, the convolution was calculated using PyTorch without using the GPU.

[0133] (1) Set evaluation indicators;

[0134] This chapter selects accuracy and recall to reflect the classification ability of the network. Accuracy refers to the proportion of correctly classified positive and negative samples out of the total number of samples. The specific calculation method is as follows:

[0135]

[0136] In the above formula, TP represents the number of positive samples that are correctly classified as positive, TN represents the number of negative samples that are correctly classified as negative, FP represents the number of positive samples that are misclassified as negative, FN represents the number of negative samples that are misclassified as positive, and the accuracy reflects the accuracy of the network in classifying the overall data.

[0137] Recall rate refers to the proportion of correctly classified positive samples out of the total number of positive samples. The specific calculation method is as follows:

[0138]

[0139] Recall reflects the network's ability to detect positive samples, but a high recall rate may lead to more false positives.

[0140] SCR is used to reflect the saliency of the target. The specific calculation of the SCR for the i-th target is as follows:

[0141]

[0142] In the above formula, m i μ represents the average gray level of the target. i and σ i These are the mean and standard deviation of the region's gray level, respectively.

[0143] The true positive rate (TPR) and false positive rate (FPR) are defined as follows:

[0144]

[0145]

[0146] In the above formula, TD represents the number of real targets detected, ND represents the total number of targets detected, FD represents the number of false targets detected, and NT represents the total number of pixels in the region;

[0147] (2) Obtain data;

[0148] Since there is currently no publicly available dataset related to the research in this invention, this invention utilizes star images captured by different monitoring telescopes (space-based and ground-based) under different imaging conditions to construct the dataset. A spatial target with a known location is selected from the star image, and an 11×11 sub-region around the target is cropped and normalized to construct the dataset. The dataset includes 3408 images, divided into training and testing sets in a 7:3 ratio. Due to the high probability of non-target regions, to avoid overemphasizing negative samples during training and to accelerate convergence, the training set contains 1428 positive samples and 956 negative samples, while the testing set contains 578 positive samples and 446 negative samples. Figure 6 and Figure 7 As shown, the positive samples include the following three cases: a) the target moves in different directions, i.e., the target extends and is distributed in different directions; b) the target has different sizes and salience; c) the image contains other stars or targets. These methods ensure sample diversity and avoid overfitting. During training, Stochastic Gradient Descent (SGD) is used for optimization, with a learning rate of 0.001, a momentum of 0.9, and a batch size of 10. The resulting loss curve is shown in the figure. Figure 8 As shown;

[0149] The performance of the proposed method was validated using four sequences of stellar images captured by ground-based space surveillance telescopes. The experiment consisted of two parts: a comparison of detection results for different targets and a comparison of detection results for the same target under different saliency levels. The data included nine typical targets from different sequences with different SCR values, with targets 7 to 9 being newly added targets. Figure 9 As shown, in order to process the algorithm quickly while preserving target information and interference, a 201×201 pixel sub-image set around the target is extracted from each frame and used to form a new sequence, denoted as Sequence 1 to Sequence 9. Detailed information about the target and background is listed in Table 2, where the information is calculated based on the real target and its surrounding area, the size of which is slightly larger than the target radius in each frame.

[0150] Table 2 Sequence Information

[0151]

[0152] To verify the detection capability of this method, it is compared with several traditional small object detection methods in recent years. Although some mainstream deep learning methods outperform traditional non-deep learning methods in many detection problems, these methods are difficult to apply directly to this problem for the following reasons: a) Due to the difference in optical systems, it is difficult to directly use prior information to obtain the correct label in the entire image, which requires a lot of time and manpower for annotation; b) Mainstream methods usually contain multiple max pooling layers, which are not suitable for targets that are too small; c) If only used for two-class recognition, the complex structure may require more training time and may not necessarily achieve better detection capabilities; d) The complex deep learning methods require more computational resources, which puts more pressure on the on-orbit system. Considering the cost of manual annotation and the feasibility in actual engineering, this chapter selects three types of methods as comparison methods: global segmentation (including max-mean, RLCM, NSM, and WSLCM), local segmentation (FKRW), and local model-driven (including SEXTRACT and MLTC). Among them, SF-CNN is implemented in Python, and the other algorithms are reproduced in C language.

[0153] To verify the impact of preprocessing on star images and their targets, a star image with strong stray light interference was selected for preprocessing. The image size was 1200×1200. The comparison images before and after processing are shown below. Figure 10 As shown;

[0154] Since the original image depth is 16 bits, to more reasonably compare their effects, both the pre- and post-processed images were stretched to the same grayscale range [121, 189]. In the contour plot, to better reflect the grayscale changes of the background, the portions of the image with grayscale values ​​higher than 200 were truncated. The image before processing contained large areas of unevenly distributed interference; the processed image showed that the interference was significantly removed and the background was close to zero. Furthermore, eight stars in different locations, affected by varying degrees of interference, and exhibiting different degrees of salience were selected for comparison. Figure 11 The image shows the images of three stars (star 1, star 4 and star 8) before and after processing, as well as their corresponding grayscale distribution. The processed image reflects more the distribution of the target above the background, i.e., the original distribution of the target. However, since the algorithm makes a rough estimate of the background, it is difficult to filter out the interference mixed in with the target. Table 3 quantitatively analyzes the processing effect, including the number of candidate points in the image and the SCR value of the target.

[0155] Table 3 Comparison of features before and after preprocessing

[0156]

[0157] It can be seen that preprocessing has little impact on image details and little interference with the original distribution of the target and its internal regions. On the other hand, due to the filtering out of large-area interference, the mis-extraction of candidate points in the bright-dark transition region is greatly reduced, which increases the processing speed of the algorithm.

[0158] Furthermore, ablation experiments were conducted under different loss functions and guidance conditions, and the results are shown in Table 4.

[0159] Table 4 Ablation experimental results under different loss functions and guidance conditions

[0160]

[0161] Compared to traditional cross-entropy loss and unguided methods, the addition of bias loss and guiding information effectively improves the network's classification accuracy and recall. Using bias loss alone instead of cross-entropy loss achieves the best recall, meaning it best ensures target detection. When bias loss and guiding information are used simultaneously, the network achieves the highest accuracy and high recall on both the training and test sets. Compared to other methods, while this approach sacrifices some probability of target detection, it effectively avoids misclassifying non-target regions as targets, significantly reducing interference targets in subsequent trajectory association subtraction.

[0162] To demonstrate the generality of this method, the above method was used to process different target sequences containing different backgrounds, shapes, sizes, and gray levels. Three typical targets (target 3, target 4, and target 6) with a wide range of SCRs were randomly selected from the different sequences. Figure 12 The corresponding saliency maps and comparisons of detection results from different methods are shown. Correctly detected space targets and stellar targets are marked with red boxes and blue circles, respectively, while yellow boxes indicate false alarms. The saliency maps for Max-mean, RLCM, NSM, WSLCM, and FKRW are the results of threshold segmentation; the MLTC column represents the final detection results; SF-CNN only provides candidate regions, and the same parameters are used for each method during detection within the same sequence.

[0163] Intuitively, due to the relatively unstable background suppression performance, false alarms and false negatives frequently occur in RLCM, NSM, WSLCM, and FKRW. Max-means, which uses target enhancement, performs well in terms of false alarms; however, when the target is in a low-contrast environment, this method may further reduce its saliency, such as... Figure 12 As shown in (B)(b), in contrast, SExtractor and MLTC focus on the distribution of targets rather than local saliency, thereby reducing false alarms and false negatives. Figure 13Four typical examples of missed targets by SExtractor and MLTC are shown, where boxes represent targets, circles represent stars, solid lines represent detected targets, and dashed lines represent missed targets. (The text then abruptly shifts to a seemingly unrelated topic: "When a target moves...") Figure 13 (A)(a)), background interference ( Figure 13 (A)(b)) or too small ( Figure 13 When (A)(c)-(d)) deviate from the ideal distribution, the above two methods also have threshold setting problems. The SF-CNN proposed in this invention follows a similar criterion, that is, it uses the intensity distribution to determine whether it is a target. The main difference between SF-CNN and the above two algorithms is that SF-CNN adopts a data-driven rather than theoretical model-driven approach. Therefore, SF-CNN is more suitable for targets with different shapes and is less affected by the background. Figure 13 (B)), Figure 14 The ROC curves for sequences 1 to 6 are shown to further demonstrate the advantages of the proposed method over the other 7 methods. As can be seen from the figure, in most cases, for the same TPR, the proposed algorithm can achieve the lowest FPR.

[0164] In addition, the statistical results of the target method are listed in Table 5;

[0165] Table 5 Detection results for different targets

[0166]

[0167] It is worth noting that when applying MLTC and SFCNN, the same parameters are shared across different sequences. In contrast, other methods determine their parameters in each sequence after balancing TPR and FPR, i.e., selecting the maximum acceptable FPR, and then determining TD / TPR at that FPR value. Overall, the highest TD and TPR across all sequences indicate that SF-CNN can stably detect targets with different SCRs and outperforms other methods. Benefiting from the diversity of the training set and the superiority of data-driven approaches, SF-CNN shows an improvement over MLTC in detection capabilities for targets with weak saliency, small FWHM, and relatively close proximity to other targets and stars (such as targets 2 and 3).

[0168] Finally, Table 6 lists the average computational cost of different methods for a single subset;

[0169] Due to preprocessing and normalization operations, SF-CNN takes more time than SExtractor and MLTC, but it has a significant advantage in processing time compared to other methods. Its computation time can be further optimized by changing the programming language.

[0170] Table 6 Computational Cost of a Single Frame Image

[0171] Max-mean PLCM NSM WSLCM FKRW SExtractor MLTC SF-CNN time consuming 141.28 538.37 1586.78 104.31 1140.12 15.55 12.89 53.85

[0172] The impact of target saliency on detection performance was further evaluated. Different exposure times were set during the acquisition of the fourth set of images to simulate the same target under different saliency conditions. The shorter the exposure time, the lower the brightness, the smaller the size, and the worse the saliency. Therefore, the algorithm needs to have stronger robustness and generalization ability.

[0173] The detection results for three targets randomly selected from sequences 7 to 9 are as follows: Figure 15 As shown in the figure, it can be observed that background clutter and noise are significantly more prevalent in the original subset with shorter exposure times compared to the subset with longer exposure times, such as... Figure 15 As shown in (A)(a), these disturbances severely reduce the saliency of the target and affect the detection performance of algorithms based on background suppression and target enhancement, such as Figure 15 As shown in (A)(b)-(g), these methods incorrectly identify many pixels as suspected targets. Although MLTC has stricter target recognition conditions, some targets that are submerged in these interferences are easily missed due to the variation in distribution. In contrast, due to reasonable preprocessing stages and candidate region search, SF-CNN can eliminate interference while preserving the original distribution of targets as much as possible. In addition, the use of irregular targets to construct the dataset makes the method have better generalization and robustness to targets with different salience.

[0174] For quantitative analysis, Table 7 lists the detection results of different methods in the fourth set of data;

[0175]

[0176] For highly saliency targets (SCR value greater than 5), these methods do not show significant differences in detection performance. However, when the target intensity is close to the background, most of the contrasting methods (including max-mean, RLCM, NSM, WSLCM, and FKRW) struggle to function properly because they are not robust to interference from similar backgrounds around the target. Even though the target is the same, its detection probability is severely affected by the target saliency. The proposed method almost always achieves the highest TPR across all test sequences, which means that SF-CNN outperforms the contrasting methods in terms of detection probability.

[0177] In summary, this invention aims to propose a data-driven space target detection method based on the distribution characteristics of space targets. By estimating the background based on local gray-level statistical information, the foreground and background can be separated simply and effectively. The processing stage can significantly reduce erroneous searches of candidate regions. In particular, unlike traditional methods based on background suppression, target enhancement, and theoretical model-driven approaches, this method transforms the global detection problem into a local two-class recognition problem. Each candidate region is classified by a trained CNN. Experiments conducted on a large number of stellar images containing complex backgrounds and targets with varying salience show that SF-CNN significantly outperforms comparative methods such as max-mean, ILCM, RLCM, NSM, WSLCM, SExtractor, and MLTC. Furthermore, this method demonstrates superiority in detection capability and the provision of additional size information. Moreover, the network structure of this method is easy to implement and generalize, possessing the potential for application in space-based detection.

[0178] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A data-driven point-like space object detection method, characterized in that, The method comprises the following steps: preprocessing the image to be detected to obtain a preprocessed image; searching the preprocessed image according to a minimum pixel-level radius of the image to obtain a candidate region; inputting the candidate region as data-driven into an SF-CNN classification model for guided classification to obtain an image prediction category; the searching the preprocessed image according to a minimum pixel-level radius of the image to obtain a candidate region specifically comprises: determining a minimum search distance between targets in the preprocessed image according to the minimum pixel-level radius of the image; determining an inflation operation region according to the minimum search distance between the targets in the preprocessed image; performing inflation search processing on the preprocessed image according to the inflation operation region to obtain a local maximum point of the image; performing set construction processing on a region where the local maximum point of the image is located to obtain the candidate region; the inputting the candidate region as data-driven into an SF-CNN classification model for guided classification to obtain an image prediction category specifically comprises: performing normalization processing on the candidate region to obtain a normalized candidate region; performing image interpolation processing on the normalized candidate region to obtain an interpolated candidate region; inputting the interpolated candidate region into the SF-CNN classification model for classification to obtain a classification result, wherein the classification result comprises a confidence value of a target and a confidence value of a non-target; determining the size of the target according to the classification result and generating a corresponding saliency map and radius map; performing redundant target elimination processing on the saliency map and the radius map by a non-maximum suppression method to obtain the image prediction category; the inputting the interpolated candidate region into the SF-CNN classification model for classification to obtain a classification result specifically comprises: inputting the interpolated candidate region into the SF-CNN classification model, wherein the SF-CNN classification model comprises a first convolutional layer, a pooling layer, a second convolutional layer and a fully connected layer; performing feature extraction processing on the interpolated candidate region based on the first convolutional layer to obtain feature information of the image; performing down-sampling operation processing on the feature information of the image based on the pooling layer to obtain low spatial dimension image feature information; performing combination processing on the low spatial dimension image feature information based on the second convolutional layer to obtain combined image feature information; based on the fully connected layer, calculating the confidence value of the interpolated candidate region according to the combined image feature information and classifying, and outputting a classification result.

2. The data-driven point-like space object detection method of claim 1, wherein, the preprocessing the image to be detected to obtain a preprocessed image specifically comprises: capturing a star map by a monitoring telescope to obtain the image to be detected; performing noise elimination processing on the image to be detected by a median filter to obtain a noise-suppressed image; performing image background value estimation processing on the noise-suppressed image to obtain a block background value; performing conversion processing on the block background value by a bilinear interpolation method to obtain a background map and a root mean square map; performing segmentation processing on the background map to obtain a foreground image, wherein the foreground image comprises a star to be detected and a target; performing binarization processing and encoding processing on the foreground image in sequence according to the root mean square map to obtain the preprocessed image.

3. The data-driven point-like space object detection method of claim 2, wherein, The step of clipping the image after noise suppression and image background value estimation of the to-be-detected image to obtain a block background value specifically comprises the following steps: iteratively clipping the image after noise suppression until the image background value meets a preset convergence range, stopping clipping, and obtaining a clipped image; judging the change of the image background value in the clipping process; when the reduction speed of the image background value in each clipping is less than a preset reduction speed, selecting the average value of the background value of the clipped image as the block background value; when the reduction speed of the image background value in each clipping is greater than the preset reduction speed, performing mode estimation on the background value of the clipped image to obtain the block background value.

4. The data-driven point-like space object detection method of claim 3, wherein, The calculation process of the mode estimation is as shown below: In the above formulae, denotes the mode estimate, denotes the median of the output of the median filter, denotes the mean.

5. The data-driven point-like space object detection method of claim 4, wherein, The expression of the non-maximum suppression method is as shown below: in the above formula, denotes the radius in the maximum value, denotes the saliency map, denotes the radius map, denotes the suspected center point, denotes a certain point on the saliency map , denotes a certain point on the radius map .

6. The data-driven point-like space object detection method of claim 5, wherein, The loss function of the second convolutional layer is a bias loss function, and the expression thereof is specifically as shown below: In the above formula, denotes the deviation loss function, denotes the nonlinear scaling function, , denotes the adjustable contribution parameter, denotes whether the th sample is of the class , denotes the feature value corresponding to the th sample, denotes the model parameter, denotes the neural network, denotes the variance of the processed th sample output by the convolutional layer, denotes the total number of samples.

Images