A space-based infrared dim small target detection method based on vector signal

Abstract

The application discloses a space-based infrared dim small target detection method based on a vector signal, in order to overcome the low timeliness and large resource consumption bottleneck of the traditional "imaging-caching-preprocessing-detection" link, skips the image caching and preprocessing link, and directly takes the DN value vector signal read out by the detector as input. Wherein, the two-dimensional or higher-dimensional features of the target are recovered from the one-dimensional vector signal through the network based on the CNN and BERT mechanisms and with the help of the Markov random field model; the network based on the self-supervised U-Net structure is used to decouple the target and background clutter features, and the multi-scale detection mechanism is combined to selectively enhance the target features to complete the detection; the network is lightened through the multi-teacher knowledge distillation architecture, and pruning and quantization are combined to obtain a lightened detection model suitable for the resource limited environment on the satellite. The application realizes the "detection and calculation integration", and significantly improves the timeliness and resource efficiency of the space-based infrared dim small target detection.

Description

Technical Field

[0001] This invention relates to on-orbit data processing and infrared weak target detection technology, and in particular to a space-based infrared weak target detection method based on vector signals. Background Technology

[0002] "Fast, accurate, and coordinated" is the core requirement for time-sensitive target detection in space-based infrared situational awareness systems. Due to the long detection range of space-based systems, the imaging area of ​​time-sensitive targets occupies only a small portion of the global image, and the radiation intensity of these targets is significantly attenuated. Therefore, space-based infrared time-sensitive target detection is essentially the detection of weak targets in space-based infrared systems. Thus, meeting the core requirements of space-based infrared sensing systems necessitates improving the timeliness of space-based infrared weak target detection methods. However, the continuously improving maneuverability of time-sensitive targets and new high spatiotemporal resolution imaging technologies make it difficult for current on-orbit processing technologies in space-based infrared sensing systems to meet real-time detection requirements. Therefore, there is an urgent need to improve the timeliness of space-based infrared weak target detection methods.

[0003] Traditional onboard processing follows the conventional "imaging-caching-preprocessing-detection" infrared detection chain, with infrared image preprocessing involving multiple steps. This leads to two major bottlenecks: 1) Excessive processing time, failing to meet the demands for rapid detection of time-sensitive targets; 2) Data caching and image preprocessing consume onboard energy and computing resources, preventing resources from being focused on detecting weak infrared targets, which in turn limits the accuracy of target detection.

[0004] Currently, space-based infrared small target detection is one of the hot topics in the field of satellite remote sensing, and related research continues to be highly productive. Scholars have conducted extensive research in both the infrared information acquisition and image processing stages (with a large number of publications annually), providing references for enhancing infrared target features in vector signals and suppressing background clutter. Existing research mainly focuses on target detection based on infrared images, with a few researchers beginning to explore using infrared image row vectors as input for real-time infrared small target detection. However, research directly using vector signals as input is rarely published.

[0005] Among mature infrared image-based methods for detecting weak targets, single-frame infrared small target detection methods offer good timeliness, are applicable to various satellite imaging modes, and do not require caching multiple frames. While multi-frame detection methods perform better in background clutter suppression, they are only applicable to staring modes and have poor timeliness. Traditional small target detection algorithms (such as filtering, local contrast, and data structure methods) are weak in suppressing complex backgrounds and strong clutter, resulting in a high false alarm rate. Deep learning algorithms, with their powerful target feature extraction capabilities and non-global dependency characteristics, exhibit significant advantages, but also face technical bottlenecks such as high algorithm complexity and high resource consumption.

[0006] Meanwhile, some researchers have begun exploring real-time infrared image preprocessing and small target detection methods using infrared row vectors as input. Addressing the impact of severe non-uniform background in spatial surveillance images on target detection and recognition performance, in 2023, Zhou et al. proposed a real-time non-uniform background correction algorithm based on local pixel vector signals. This algorithm breaks through the dependence on information from the entire image, innovatively using single-row local vector signals for real-time processing, achieving ultra-low latency infrared image preprocessing at the microsecond level, and completing efficient hardware implementation through FPGA pipeline design. To improve the latency of single-frame infrared small target detection methods, Dali et al. proposed the 1-D Bidirectional Vector Feature Measure (1-DBVFM) method. This algorithm uses row data from the infrared image as input, suppressing background clutter and enhancing infrared small targets through manually designed 1-D bidirectional vector operators and 1-D bidirectional vector gradient operators, respectively, and finally outputs the target position through adaptive global threshold segmentation. However, the 1-DBVFM method still uses mature infrared images as input and does not consider the influence of image noise and non-uniformity.

[0007] Existing space-based infrared weak target detection methods based on infrared images rely on traditional imaging detection chains (imaging-buffering-preprocessing-detection), thus failing to meet the high-timeliness processing requirements of hypersonic time-sensitive target detection and high-throughput data in wide-swath detection systems. While image row vector-based target detection methods improve the real-time performance of on-orbit detection to some extent, they still essentially use mature infrared images as input and do not consider the effects of image noise and non-uniformity. Therefore, the timeliness of such methods still falls short of the high-timeliness processing requirements of hypersonic time-sensitive target detection and high-throughput data in wide-swath detection systems.

[0008] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a space-based infrared weak target detection method based on vector signals.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A space-based infrared detection method for weak targets based on vector signals includes the following steps:

[0012] S1. Target Multidimensional Feature Extraction: Using the digital numerical (DN) vector signal from the row readout after imaging by the infrared detector as input, a target multidimensional feature extraction network is constructed and trained to recover the two-dimensional or higher-dimensional features of the target from the one-dimensional vector signal. The feature extraction network adopts the idea of ​​local caching and dynamic two-dimensional feature splicing, and enhances the inter-row feature relationship of the same object by using the inter-row signal correlation model.

[0013] S2. Coupling Feature Decoupling Enhancement and Target Detection: Based on the recovered high-dimensional target features, a coupling feature decoupling enhancement and target detection network is constructed and trained to decouple target features from background clutter or interference features, and selectively enhance target features to complete the detection of weak infrared targets; the network adopts a self-supervised learning framework and combines a multi-scale detection mechanism to capture the feature differences between the target and the background at different scales;

[0014] S3. Network Lightweighting and Deployment: Based on the network trained in steps S1 and S2, a multi-teacher knowledge distillation architecture is used to lightweight it, resulting in a lightweight detection network suitable for on-board resource-constrained environments. Further model pruning and quantization optimization are then performed.

[0015] A computer program product includes a computer program that, when executed by a processor, implements the aforementioned vector signal-based space-based infrared weak target detection method.

[0016] The present invention has the following beneficial effects:

[0017] This invention proposes a vector signal-based integrated detection method for space-based infrared weak targets, which significantly improves the timeliness of space-based infrared weak target detection and meets the real-time processing requirements of high-throughput data in hypersonic time-sensitive target detection and wide-swath imaging systems. Unlike existing technologies, this invention skips the traditional "readout-caching-preprocessing-global detection" process and directly uses the digital numerical (DN) row vector signal read out line by line by the infrared detector as the processing object. Target detection begins simultaneously with the imaging signal readout, achieving "integrated detection and calculation." This significantly shortens the data processing link, reduces the resource consumption of on-board data caching and preprocessing, and provides a feasible technical path for achieving "fast, accurate, and coordinated" detection of space-based infrared weak targets.

[0018] Specifically, this invention proposes optimized solutions to two core problems in signal-level detection: First, for the problem of how to recover multi-dimensional features of a target from a one-dimensional vector signal, a high-dimensional feature recovery network based on a convolutional neural network (CNN) and bidirectional encoder representation (BERT) mechanism is constructed. This network is combined with a Markov random field (MRF) model to characterize and enhance the correlation of inter-row vector signals, thereby effectively reconstructing the two-dimensional or higher-dimensional features of the target from the serialized row vectors. Second, for the problem of target features being mixed with background clutter and noise, making effective enhancement difficult, a coupled feature decoupling enhancement and target detection module is designed. This module decouples target and interference features through a self-supervised U-Net network and constructs a multi-scale detection head using the idea of ​​a dynamic CNN combined with a YOLO network to selectively enhance target features. Finally, it is connected in series with a feature extraction network to achieve end-to-end detection of weak targets.

[0019] To address the severe constraints of limited storage and computing resources on space-based platforms, this invention further introduces a lightweight dual-teacher knowledge distillation module. This module uses a CNN+BERT feature extraction network and a U-Net+CNN+FPN detection network as teacher models. Through knowledge distillation, its capabilities are transferred to student feature extraction models composed of lightweight CNNs and Long Short-Term Memory (LSTM) networks, and student detection models composed of lightweight CNNs and lightweight target detection networks (such as the SOCDet architecture). Furthermore, the lightweight networks can be pruned and quantized to further reduce model size and computational overhead, ensuring the entire method can operate stably and efficiently in the limited resource environment of spacecraft. This provides a high-efficiency, low-consumption real-time detection solution for weak targets in space-based infrared situational awareness systems.

[0020] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the space-based infrared weak target detection method based on vector signals according to the present invention.

[0022] Figure 2 This is a high-dimensional feature recovery network diagram based on the CNN+BERT mechanism in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of a weak target detection method for infrared vector signals based on self-supervised learning, according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the lightweight module for multi-teacher knowledge distillation in an embodiment of the present invention.

[0025] Figure 5This is a flowchart of the space-based infrared weak target detection method based on vector signals according to an embodiment of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0027] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0028] This invention aims to address the bottlenecks of long and slow detection processes in traditional space-based infrared weak target detection. It proposes an integrated detection method that skips image caching and preprocessing steps and directly processes the detector's readout vector signal. By constructing a signal-oriented feature recovery, decoupling enhancement, and lightweight distillation network, it achieves instant detection and calculation while significantly improving detection timeliness and reducing onboard resource consumption, effectively meeting the needs of rapid detection and high-throughput data processing for hypersonic time-sensitive targets.

[0029] See Figures 1 to 5 This invention provides a method for integrated detection of weak targets using space-based infrared sensors based on vector signals, comprising the following steps:

[0030] Step S1, Target Multidimensional Feature Extraction: Using the digital numerical (DN) vector signal read out after imaging by the infrared detector as input, construct and train a target multidimensional feature extraction network to recover the two-dimensional or higher-dimensional features of the target from the one-dimensional vector signal; the feature extraction network adopts the idea of ​​local caching and dynamic two-dimensional feature splicing, and enhances the inter-row feature relationship of the same object by means of the inter-row signal correlation model.

[0031] In some embodiments, step S1 specifically includes: constructing a high-dimensional feature recovery network based on a convolutional neural network (CNN) and bidirectional encoder representation (BERT) mechanism, wherein the network includes multiple BERT feature modules, inter-line feature concatenation modules, and a dynamic CNN high-dimensional feature recovery network connected in sequence; inputting the vector signal sequence read out line by line by the detector into the high-dimensional feature recovery network, performing word embedding and position embedding aggregation on the vector signal through the BERT feature modules to extract features of the same object within the vector signal; using the inter-line feature concatenation module, employing an attention mechanism to understand the relationship between adjacent inter-line features, and combining it with an inter-line correlation model based on a Markov random field (MRF), which enhances the inter-line feature relationship of the same object by calculating the difference between adjacent row vectors and based on a probability model, and outputting a static high-dimensional feature matrix; the dynamic CNN high-dimensional feature recovery network receiving the static high-dimensional feature matrix and outputting a dynamic high-dimensional feature matrix to recover the multi-dimensional features of the target; using the vector signal and its corresponding zero-level infrared image as training data and labels, constructing a loss function in conjunction with the MRF model, and training the high-dimensional feature recovery network.

[0032] In some embodiments, the inter-row correlation model is a Markov random field (MRF) model, used to characterize and enhance the feature correlation of the same objects in the inter-row vector signal.

[0033] Step S2, Coupling Feature Decoupling Enhancement and Target Detection: Based on the recovered high-dimensional target features, a coupling feature decoupling enhancement and target detection network is constructed and trained to decouple target features from background clutter or interference features, and selectively enhance target features to complete the detection of small infrared targets; the network adopts a self-supervised learning framework and combines a multi-scale detection mechanism to capture the feature differences between the target and the background at different scales.

[0034] In some embodiments, step S2 specifically includes: constructing a feature decoupling network based on a U-Net structure, and introducing a proxy attention mechanism in a specific layer of the network, wherein the proxy attention mechanism enhances the ability of the traditional attention module to extract global features by introducing additional proxy tokens; pre-training the feature decoupling network using a self-supervised learning method with a zero-level infrared image as input and a pre-processed mature infrared image as a label; constructing a target detection head network, wherein the network includes a dynamic CNN network and a multi-scale detection head based on a feature pyramid network (FPN); the dynamic CNN network is used to adjust the feature dimensions to match the input, and the multi-scale detection head is used to selectively enhance and detect weak infrared targets at different scales; sequentially concatenating the trained target multi-dimensional feature extraction network, the feature decoupling network, and the target detection head network to construct an end-to-end weak infrared target detection network; constructing a loss function based on the physical distribution characteristics of the target and the differences in the spatial and channel dimensions of the feature map, and fine-tuning the end-to-end detection network as a whole to enhance the target detection head module's ability to enhance and detect the features of weak infrared targets.

[0035] Step S3, Network Lightweighting and Deployment: Based on the network trained in steps S1 and S2, a multi-teacher knowledge distillation architecture is used to lightweight it, resulting in a lightweight detection network suitable for on-board resource-constrained environments. Further model pruning and quantization optimization are then performed.

[0036] In some embodiments, the multi-teacher knowledge distillation architecture in step S3 specifically includes: using the target multi-dimensional feature extraction network trained in step S1 as the first teacher model, and using the coupled feature decoupling enhancement and target detection network trained in step S2 as the second teacher model; constructing a lightweight student model, wherein the student model includes a lightweight CNN and Long Short-Term Memory (LSTM) network combination for feature extraction, and a lightweight CNN and lightweight target detection network combination for target detection; designing a comprehensive distillation loss function, wherein the loss function includes at least: the feature map loss between the feature extraction outputs of the first teacher model and the student model, The loss includes the feature map loss between the target detection outputs of the second teacher model and the student model, the task loss of the student model on the original detection task, and the feature loss between the feature extraction output of the student model and the real feature data. The feature map loss is measured by calculating the mean square error of the values ​​at all positions in the channel, height, and width dimensions of the two feature maps. The comprehensive loss function is a weighted sum of the above types of losses, and the contribution of each type of loss to the total loss is balanced by hyperparameters. The student model is trained using the comprehensive loss function, and knowledge is distilled from the first teacher model and the second teacher model to the student model to obtain a lightweight detection network.

[0037] In some embodiments, the comprehensive distillation loss function is a weighted sum of the feature map loss, the task loss, and the feature loss, wherein the various losses are balanced by hyperparameters.

[0038] In some embodiments, the lightweight design specifically includes: employing depthwise separable convolution, pointwise convolution, and fully connected layer parameterization methods in the dynamic CNN part of the student model to reduce the number of parameters; the lightweight object detection network is designed based on a lightweight object detector architecture.

[0039] In some embodiments, step S3 further includes: after obtaining the lightweight detection network, further evaluating and implementing network pruning and quantization operations to reduce model size and computational resource consumption, so as to adapt it to the constraints of the space-based platform.

[0040] In some embodiments, the space-based infrared weak target detection method based on vector signals skips the traditional "read-buffering-preprocessing-global detection" process and directly realizes the detection of infrared weak targets during the process of reading out vector signals from the infrared detector.

[0041] This invention proposes a space-based infrared weak target detection method based on vector signals. By skipping the lengthy traditional "imaging-caching-preprocessing-detection" process, it directly uses the one-dimensional DN vector signal read from the infrared detector as the starting point. It constructs a complete integrated detection and calculation process, including multi-dimensional target feature extraction, coupled feature decoupling enhancement and detection, and dual-teacher distillation lightweighting. This fundamentally solves the timeliness bottleneck caused by process delays and resource dispersion in existing methods. Furthermore, it directly recovers and enhances target features from the signal using CNN+BERT and MRF models, and then achieves accurate feature decoupling and enhancement through a self-supervised U-Net and multi-scale detection head. Multi-teacher distillation and subsequent pruning quantization are also used to adapt the network to the stringent resource constraints of spaceborne applications. This invention significantly shortens processing time and reduces onboard energy consumption while ensuring high-timeliness and high-precision detection capabilities for hypersonic time-sensitive targets and high-throughput data from wide-swath detection, achieving the core requirement of "fast, accurate, and collaborative" processing.

[0042] The following further describes specific embodiments of the present invention and examples of its algorithm implementation.

[0043] A space-based infrared detection method for weak targets based on vector signals comprises three core components: multidimensional target feature extraction, coupled feature decoupling enhancement and target detection, and lightweight network deployment. First, to address the core issue of recovering multidimensional target features from one-dimensional vector signals, a high-dimensional feature recovery network based on a convolutional neural network (CNN) and bidirectional encoder representation (BERT) mechanism is constructed. This network extracts features from the vector signal using the BERT module and leverages a Markov random field (MRF) model to characterize and enhance the correlation between inter-row vector signals, thereby effectively recovering two-dimensional or higher-dimensional target features from sequential row vectors.

[0044] To decouple target features from background clutter and other interference features, and to selectively enhance target features for detection, a coupled feature decoupling enhancement and target detection module was constructed. This module employs a U-Net-based feature decoupling network, incorporating a surrogate attention mechanism to enhance global feature extraction capabilities. This feature decoupling network is pre-trained using self-supervised learning with zero-level infrared and mature infrared images. Subsequently, it is combined with a target detection head that incorporates a dynamic CNN network and a feature pyramid network (FPN) based on multi-scale detection (integrating ideas from YOLO networks), and then concatenated with the aforementioned multi-dimensional target feature extraction network to form an end-to-end infrared weak target detection network. Fine-tuning is then used to optimize detection performance.

[0045] To adapt to the severely limited storage and computing resources of space-based platforms, a lightweight design was implemented for the aforementioned detection network. A dual-teacher knowledge distillation architecture was adopted, using the aforementioned CNN+BERT-based feature extraction network and the U-Net+CNN+FPN-based detection network as teacher models. Their knowledge was distilled into a feature extraction student model composed of a lightweight CNN and a Long Short-Term Memory (LSTM) network, and a detection student model composed of a lightweight CNN and a lightweight object detection network (such as the SOCDet architecture). A lightweight student network was obtained through training using a comprehensive distillation loss function. To further compress the model, pruning and quantization operations were performed on this lightweight network, ultimately resulting in a high-efficiency, low-resource-consumption infrared weak target detection scheme suitable for space-based deployment.

[0046] This innovative method breaks away from the traditional "imaging-caching-preprocessing-detection" chain, directly using the digital numerical (DN) row vector signal read out after the infrared detector images as the processing object. Target detection begins at the signal readout stage, realizing "integrated detection and computation". This significantly shortens the data processing flow, reduces the resource consumption caused by on-board data caching and preprocessing, and ultimately improves the timeliness of space-based infrared weak target detection.

[0047] The following describes the detailed processing flow and core module structure of this method. (See attached document.) Figure 5 The detection method includes a target multidimensional feature extraction module, a coupled feature decoupling enhancement and target detection module, and a dual-teacher distillation lightweight module.

[0048] 1. Target Multidimensional Feature Extraction Module

[0049] The original three-dimensional target is reduced to a two-dimensional target after infrared focal plane imaging. However, the detector outputs only a one-dimensional vector signal after analog-to-digital conversion, causing the two-dimensional features of the target to be further decomposed into a one-dimensional space. This makes the core scientific problem of infrared weak target detection based on vector signals: how to recover the multi-dimensional features of the target from the 1-D vector signal. Traditional infrared image small target detection methods directly construct target features in two-dimensional space, but this scientific problem is unavoidable when processing vector signals. Therefore, this invention constructs a target multi-dimensional feature extraction module to recover the multi-dimensional features of the target from the one-dimensional vector signal.

[0050] This invention proposes to construct a high-dimensional feature extraction network for infrared vector signals using a combination of local caching and dynamic two-dimensional feature concatenation, and to construct a high-dimensional feature recovery network using a CNN+BERT mechanism. The network structure is as follows: Figure 2 As shown. Then, a Markov random field (MRF) model is used to characterize the correlation of inter-row vector signals, enhancing the inter-row feature relationships of the same objects. The MRF model is as follows:

[0051]

[0052] Wherein represents This represents the probability of inter-row correlation. Indicates the first The row interval centered on the vector. For the interval radius hyperparameter, These are the relevant hyperparameters.

[0053] The operation steps of the 1-D vector high-dimensional feature extraction module based on the CNN+BERT mechanism are as follows:

[0054] (1) Constructing if A series of interconnected BERT feature modules ( The value is slightly larger than the size of small infrared targets, such as Each BERT model first aggregates and outputs 1-D features of the same object (target or background) within the vector signal through word embedding and position embedding;

[0055] (2) Construct an inter-line feature concatenation module, use BERT mid-segment embedding attention to understand the relationship between 1-D features of adjacent lines, then combine MRF to extract inter-line features of the same object, and finally output a static high-dimensional matrix. ;

[0056] (3) Construct a dynamic CNN high-dimensional feature recovery network with an output dimension of Dynamic high-dimensional feature matrix ;

[0057] (4) Using the vector signals and zero-level images in the dataset as training data, input the vector signals row by row. , The number of rows in the image is the batch number, and the corresponding zero-level image is the number of rows. For tags;

[0058] (5) Design based on MRF model and The loss function between them is used to train a high-dimensional feature extraction network for vector signals;

[0059] (6) A well-trained network can represent Then, the interline feature stitching module and the BERT feature module are extracted for fine-tuning during subsequent end-to-end vector signal infrared target detection.

[0060] 2. Coupling Feature Decoupling Enhancement and Target Detection Module

[0061] After recovering the high-dimensional features of the target, signal-level infrared target detection still faces two problems: how to decouple the target-clutter features and how to selectively enhance the target features. This invention constructs a hybrid feature system that simultaneously enhances target features by decoupling and enhancing coupled features and decoupling the target detection module, thereby achieving infrared small target detection.

[0062] This invention employs self-supervised learning and a U-Net architecture to decouple target features from interference features. Then, a dynamic CNN is constructed in conjunction with a YOLO network to selectively enhance target features, achieving infrared detection of small targets. The FPN multi-scale detection head in the YOLO architecture captures different features of the target and background at different scales, improving the detection accuracy for small targets.

[0063] Figure 3 The following is a flowchart illustrating the target-interference hybrid feature decoupling mechanism and the method for detecting weak targets using infrared vector signals: First, a feature decoupling network is constructed using self-supervised learning, with zero-level images as samples and mature infrared images as labels, to pre-train a U-Net-structured feature decoupling network. Then, a 1-D vector high-dimensional feature extraction network based on CNN+BERT and the feature decoupling network are combined to construct and fine-tune a vector signal-based infrared weak target detection system to achieve end-to-end detection of weak infrared targets based on vector signals. The specific method flow is as follows:

[0064] (1) Construct a feature decoupling network based on the U-Net structure. Before the last downsampling, after the first upsampling, and in the residual blocks of the intermediate layers, a proxy attention mechanism is used to extract global feature information at different preprocessing stages. The proxy attention module can be represented as a quadruple (Q, A, K, V), which introduces an additional set of agent tokens A into the traditional attention module.

[0065] (2) Feature decoupling network training. Using the zero-level image and the corresponding infrared image as labels, a loss function was constructed based on the mathematical model of infrared image preprocessing for self-supervised training. The trained infrared feature decoupling network awaits further fine-tuning.

[0066] (3) Construct the target detection head network. The FPN target detection head is constructed by combining the idea of ​​dynamic CNN with YOLO. First, the infrared feature decoupling network dimension is adjusted by the dynamic CNN network to match the number of input vector signals. Then, the YOLO network selectively enhances the infrared target.

[0067] (4) The high-dimensional feature extraction network of vector signals, the feature decoupling network, and the target detection head are sequentially concatenated to construct an end-to-end infrared weak target detection network based on vector signals. A loss function is constructed by combining the physical distribution characteristics of the target to selectively enhance the target features. Further fine-tuning is performed to improve the target detection head module's ability to enhance the features of infrared weak targets, and finally end-to-end infrared weak target detection is achieved.

[0068] 3. Lightweight Module for Dual-Teacher Distillation

[0069] Due to the limited resources of the onboard processing platform, a lightweight design is needed for the infrared vector signal weak target detection method based on self-supervised learning to ensure its normal operation on the satellite with severely limited computing resources. The infrared weak target detection method based on vector signals proposed in the first two steps consists of a high-dimensional feature recovery module and a coupled feature decoupling enhancement and target detection module. Therefore, a multi-teacher distillation architecture is proposed to lightweight the above network.

[0070] Multi-teacher distillation is a method that improves the performance of a student network by leveraging the collaborative efforts of multiple teacher networks, enabling better guidance for the student network in learning richer and more diverse knowledge. The high-dimensional feature extraction network in the teacher network uses the CNN+BERT architecture mentioned above, while the weak object detection network uses the U-Net+CNN+FPN architecture. In the student network, the high-dimensional feature extraction network is selected as a dynamically lightweight CNN+LSTM model, and the weak object detection network is selected as a dynamically lightweight CNN+lightweight object detection network. For example... Figure 4As shown, the high-level feature extraction network and the weak target detection network in the teacher network are respectively represented as follows: and The high-level feature extraction network and weak target detection network of the student network are respectively represented as follows: and The specific method is as follows:

[0071] (1) Construct a dual-teacher distillation architecture. In the teacher model... It converts the input one-dimensional signal into a two-dimensional feature map. (In the student model...) Through learning Intermediate representation of the output (e.g.) The last layer of features (or their activation values) is used to understand the row correlation in the input vector signal.

[0072] (2) Design of distillation loss function. For and The output matrix is ​​represented by . Loss between feature maps (LOSS); for and The output matrix is ​​adopted. To represent the loss of the feature map; for And feature-level data truth values, using The loss is expressed as follows. The above losses can be uniformly represented by the following formula: (3)

[0073] in, and These are feature maps F1 and F2 in the channel. ,high ,width The value at the location. The resulting loss reflects the overall difference between the two feature maps in the spatial and channel dimensions. It measures the difference between the student model and the teacher model in the feature space. The loss of the student model on the original task consists of classification loss, localization loss, and confidence loss. Therefore, the comprehensive loss function is shown in the following equation:

[0074] (4)

[0075] in, , , and It is a hyperparameter used to balance various losses.

[0076] (3) Lightweight design of dynamic CNN. The number of parameters in the feature concatenation module and dynamic CNN module is reduced by using methods such as depthwise convolution, pointwise convolution and parameterization of fully connected layer matrices. Then, a lightweight object detection head is designed based on the lightweight object detection network SOCDet.

[0077] (4) Dual-teacher distillation architecture training. After training, the student model is retained as a lightweight detection network for infrared weak targets based on space-based vector signals.

[0078] (5) Redundancy characterization of the lightweight detection network for weak infrared targets based on space-based vector signals. The pruning and quantization properties of the lightweight detection network for weak infrared targets based on space-based vector signals are evaluated, and the pruning and quantization of the lightweight detection network for weak infrared targets based on space-based vector signals are further improved.

[0079] In one example, a semi-simulated dataset of weak infrared targets using space-based vector signals is constructed. This dataset is built using a process of "background image collection - weak infrared target generation - target background overlay - background degradation - row readout," with image sizes consistent with the original imaging detector array. The dataset includes zero-level images and corresponding pixel-level labels. The test set and virtual set are divided at a 1:10 ratio. During training, signal-level row vectors and corresponding row labels are extracted row by row from the training images and input into the model. During testing, signal-level row vectors and corresponding row labels are extracted row by row from the test images and input into the model to verify the detection results. The row input length is set to match the number of columns in the detector array, the batch size is 8, and the training epochs are 200. The initial learning rate is 0.01, dynamically adjusted; if the loss function does not decrease after 10 consecutive epochs, the learning rate is reduced to 0.5 times the original. The selected optimizer is Adam.

[0080] The specific training steps are as follows: 1. Input the dataset and labeled data into the network in batches for forward propagation. 2. The network outputs the prediction results and calculates the loss value with the labels. 3. The network performs backpropagation based on the loss value and updates the weight parameters through the Adam optimizer. 4. Repeat the above process continuously, and the loss value continuously decreases until the required number of training rounds is reached.

[0081] In summary, the proposed space-based infrared weak target detection method based on vector signals uses detector vector signals as the processing object. Addressing the scientific problems of high-dimensional target feature recovery and decoupling and selective enhancement of hybrid features, it skips the "caching-preprocessing" stage to achieve instant detection and computation. This invention is expected to significantly improve the timeliness of infrared weak target detection, enhance the "detection-computation integration," and provide a high-timeliness sensing solution for space-based infrared situational awareness systems.

[0082] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0083] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0084] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0085] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0086] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0087] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0088] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0089] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0091] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0092] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0093] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0094] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A space-based infrared detection method for weak targets based on vector signals, characterized in that, Includes the following steps: S1. Target multidimensional feature extraction: Using the digital value (DN) vector signal read out after imaging by the infrared detector as input, construct and train a target multidimensional feature extraction network to recover the two-dimensional or higher-dimensional features of the target from the one-dimensional vector signal. The feature extraction network employs a combination of local caching and dynamic two-dimensional feature concatenation, and leverages an inter-row signal correlation model to enhance the inter-row feature relationships of the same objects. Step S1 specifically includes: constructing a high-dimensional feature recovery network based on a convolutional neural network (CNN) and bidirectional encoder representation (BERT) mechanism; inputting the vector signal sequence read out line by line from the detector into the high-dimensional feature recovery network to extract features of the same objects within the vector signals; using the inter-row feature concatenation module, employing an attention mechanism to understand the relationship between adjacent inter-row features, and combining it with an inter-row correlation model based on a Markov random field (MRF). This model enhances the inter-row feature relationships of the same objects by calculating the differences between adjacent row vectors and based on a probability model, outputting a static high-dimensional feature matrix; the dynamic CNN high-dimensional feature recovery network receives the static high-dimensional feature matrix and outputs a dynamic high-dimensional feature matrix to recover the multi-dimensional features of the target; using the vector signals and their corresponding zero-level infrared images as training data and labels, and combining the MRF model to construct a loss function, the high-dimensional feature recovery network is trained. S2. Coupling Feature Decoupling Enhancement and Target Detection: Based on the recovered high-dimensional target features, a coupling feature decoupling enhancement and target detection network is constructed and trained to decouple target features from background clutter or interference features, and selectively enhance target features to complete the detection of weak infrared targets. The network adopts a self-supervised learning framework and combines a multi-scale detection mechanism to capture the feature differences between the target and the background at different scales. Step S2 specifically includes: constructing a feature decoupling network based on a U-Net structure, and introducing a proxy attention mechanism in a specific layer of the network. The proxy attention mechanism enhances the ability of the traditional attention module to extract global features by introducing additional proxy tokens; constructing a target detection network. The network comprises a dynamic CNN network and a multi-scale detection head based on a feature pyramid network (FPN). The dynamic CNN network is used to adjust the feature dimensions to match the input, and the multi-scale detection head is used to selectively enhance and detect infrared weak targets at different scales. The trained target multi-dimensional feature extraction network, the feature decoupling network, and the target detection head network are sequentially concatenated to construct an end-to-end infrared weak target detection network. A loss function is constructed based on the target's physical distribution characteristics and the differences in feature maps in spatial and channel dimensions to fine-tune the end-to-end detection network, thereby enhancing the target detection head module's ability to enhance and detect infrared weak target features. S3. Network Lightweighting and Deployment: Based on the network trained in steps S1 and S2, a multi-teacher knowledge distillation architecture is used to lightweight it, resulting in a lightweight detection network suitable for on-board resource-constrained environments. Further model pruning and quantization optimization are then performed.

2. The space-based infrared weak target detection method based on vector signals as described in claim 1, characterized in that, Step S1 specifically also includes: The high-dimensional feature recovery network includes multiple BERT feature modules, interline feature splicing modules, and a dynamic CNN high-dimensional feature recovery network connected in sequence. The BERT feature module performs word embedding and position embedding aggregation on the vector signal.

3. The space-based infrared weak target detection method based on vector signals as described in claim 2, characterized in that, The inter-row correlation model is a Markov random field (MRF) model, used to characterize and enhance the feature correlation of the same objects in the inter-row vector signal.

4. The space-based infrared weak target detection method based on vector signals as described in claim 1, characterized in that, Step S2 also includes: Using zero-level infrared images as input and preprocessed mature infrared images as labels, the feature decoupling network is pre-trained using a self-supervised learning approach.

5. The space-based infrared weak target detection method based on vector signals as described in claim 1, characterized in that, The multi-teacher knowledge distillation architecture described in step S3 specifically includes: The target multidimensional feature extraction network trained in step S1 is used as the first teacher model, and the coupled feature decoupling enhancement and target detection network trained in step S2 is used as the second teacher model. A lightweight student model is constructed, comprising a lightweight CNN and a Long Short-Term Memory (LSTM) network combination for feature extraction, and a lightweight CNN and a lightweight object detection network combination for object detection. Design a comprehensive distillation loss function, which includes at least: feature map loss between the feature extraction outputs of the first teacher model and the student model, feature map loss between the target detection outputs of the second teacher model and the student model, task loss of the student model on the original detection task, and feature loss between the feature extraction output of the student model and the real feature data. The feature map loss is measured by calculating the mean square error of the values ​​at all positions in the channel, height, and width dimensions of the two feature maps; the comprehensive loss function is a weighted sum of the above types of losses, and the contribution of each type of loss to the total loss is balanced by hyperparameters. The student model is trained using the comprehensive loss function, and knowledge is distilled from the first teacher model and the second teacher model into the student model to obtain a lightweight detection network.

6. The space-based infrared weak target detection method based on vector signals as described in claim 5, characterized in that, The comprehensive distillation loss function is a weighted sum of the feature map loss, the task loss, and the feature loss, wherein the various losses are balanced through hyperparameters.

7. The space-based infrared weak target detection method based on vector signals as described in claim 5, characterized in that, The lightweight design specifically includes: using depthwise separable convolution, pointwise convolution, and fully connected layer parameterization methods in the dynamic CNN part of the student model to reduce the number of parameters; the lightweight object detection network is designed based on a lightweight object detector architecture.

8. The integrated detection method for space-based infrared weak targets based on vector signals as described in any one of claims 5 to 7, characterized in that, Step S3 also includes: after obtaining the lightweight detection network, further evaluating and implementing network pruning and quantization operations to reduce model size and computational resource consumption, so as to adapt it to the constraints of the space-based platform.

9. The space-based infrared weak target detection method based on vector signals as described in claim 1, characterized in that, The method directly detects weak infrared targets during the process of reading out vector signals from the infrared detector.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the integrated detection method for space-based infrared weak targets based on vector signals as described in any one of claims 1 to 9.

Images