Single photon target identification method and device based on semantic knowledge base

Abstract

The invention discloses a single photon target recognition method and device based on a semantic knowledge base, and belongs to the technical field of data recognition. The method comprises the following steps: firstly, training a neural network model adopting a probability step variational auto-encoder structure, mapping a single photon time histogram into a mean vector and a standard deviation vector, and constructing an initial semantic knowledge base according to the mean vector and the standard deviation vector; during recognition, semantic features of a to-be-detected target are extracted, and whether the semantic features of the target are known relative to the semantic knowledge base or not is judged through unknown detection. If yes, semantic matching is carried out to output categories; and if the information is unknown, dynamically adding the characteristics of the information into the knowledge base to realize online learning and self-updating of the system. According to the method, the robustness of noise signals is enhanced through probabilistic modeling, open world recognition is realized through unknown detection, and the problems that a traditional method misjudges an unknown target and is difficult to adapt to a new environment are solved.

Description

Technical Field

[0001] This invention belongs to the field of data recognition technology, specifically relating to a method and apparatus for single-photon target recognition based on a semantic knowledge base. Background Technology

[0002] Target recognition is a fundamental task in artificial intelligence, widely applied in fields such as autonomous driving navigation, air defense, and remote sensing. In existing single-pixel single-photon detection-based imaging-free target recognition methods, laser echo signals are detected by a single-pixel single-photon detector. A time-to-digital converter records the photon flight time and converts it into a time histogram. Then, deep learning models (such as U-Net, ResNet, and MLP) are used for data analysis and recognition.

[0003] However, the aforementioned target recognition method employs a closed-set recognition model, meaning it can only identify a limited number of known categories predefined during the training phase. In practical applications, if an unknown target appears that was not seen during training, the system will force it to be classified into a known category, leading to severe recognition failures. Furthermore, when noise levels are high or the acquisition time is short, the number of received photons decreases, the effective information in the time histogram is sharply reduced, and the recognition accuracy drops significantly. Summary of the Invention

[0004] Based on this, the present invention provides a single-photon target recognition method and apparatus based on a semantic knowledge base, aiming to solve the problems of poor adaptability of traditional single-photon target recognition technology in open dynamic environments and insufficient robustness in the face of harsh signal conditions.

[0005] To achieve the above objectives, this invention provides a single-photon target recognition method based on a semantic knowledge base, the method comprising the following steps:

[0006] S1: Train a neural network model to construct a semantic knowledge base. The neural network model adopts a probabilistic step variational autoencoder structure to extract semantic features from the time histogram, and models the semantic features of different categories as multivariate Gaussian distributions. The average value of the semantic features of the same category is taken as the category semantic center and stored in the semantic knowledge base.

[0007] S2: Obtain the time histogram of the target.

[0008] S3: Input the time histogram of the target into the neural network model to extract the semantic features of the target.

[0009] S4: Perform unknown detection on the semantic features of the target to determine whether the target is known relative to the semantic knowledge base. If yes, proceed to step S5. If no, proceed to step S6.

[0010] S5: Perform semantic matching on the semantic features of the target relative to the semantic knowledge base, and output the matched known category as the recognition result.

[0011] S6: Take the average of the semantic features of the target as the semantic center of the new category, add it to the semantic knowledge base to update the semantic knowledge base, and output the new category as the recognition result.

[0012] Furthermore, the neural network model includes an encoder and a decoder.

[0013] The encoder encodes the temporal histogram into a latent space to obtain a mean vector and a standard deviation vector, and outputs them to the decoder. The encoder takes the mean vector as the semantic feature.

[0014] The encoder is also used to perform one-hot encoding of the category labels, which is mapped to the latent space by the fully connected layer of the encoder to obtain the predicted semantic center, and output to the decoder.

[0015] The decoder is used to sample the latent representation distribution based on the mean vector and the standard deviation vector, and to calculate the reconstructed histogram output based on the latent representation distribution. The decoder is also used to calculate the similarity between the semantic features and the predicted semantic centers, and to take the category label corresponding to the maximum similarity as the predicted category label output.

[0016] The predicted semantic center corresponds one-to-one with the category label. The latent representation distribution conforms to a Gaussian distribution centered at the mean vector and with the standard deviation vector as the radius. The reconstructed histogram is the reconstructed time histogram.

[0017] Furthermore, the neural network model also includes a loss function. The loss function includes divergence loss, reconstruction loss, and classification loss. The reconstruction loss characterizes the difference between the temporal histogram and the reconstructed histogram. The divergence loss characterizes the difference between the latent representation distribution and the reference distribution. The classification loss characterizes the similarity or difference between the output predicted class label and the input class label. The reference distribution conforms to a Gaussian distribution centered at the predicted semantic center and with the standard deviation vector as its radius.

[0018] Preferably, the specific steps of S1 are as follows:

[0019] S11: Collect a training sample set, which includes several sets of sample data. Each set of sample data includes the time histogram and the category label. S12: Input the training sample set into the neural network model. Minimize the loss function through iterative evolution, making the output latent representation distribution as close as possible to the reference distribution, making the output reconstructed histogram as close as possible to the input time histogram, and making the output predicted category label match the input category label. S13: After training, output the category semantic center to the semantic knowledge base to construct the semantic knowledge base.

[0020] Preferably, the specific steps of S4 are as follows:

[0021] S41: Calculate the likelihood function of the semantic features of the target relative to each known category in the semantic knowledge base, to characterize the probability that the target belongs to each of the known categories. The likelihood function is specifically expressed as: , In the formula, Let be the likelihood function. The semantic features of the target. The semantic center of the m-th known category is given. The category semantic center of the m-th known category The first component, The category semantic center of the m-th known category The d-th component, semantic features of the target The first component, semantic features of the target The d-th component, For the m-th known category, a multivariate Gaussian distribution is given. Let be the vector of integral variables in the latent space. The vector of integral variables in the latent space The first component, The vector of integral variables in the latent space The d-th component.

[0022] S42: If the maximum value of each likelihood function is greater than the likelihood threshold, then the target is determined to be known relative to the semantic knowledge base. Otherwise, the target is determined to be unknown relative to the semantic knowledge base.

[0023] Preferably, the specific steps of S5 are as follows:

[0024] Calculate the matching function of the semantic features of the target relative to each known category in the semantic knowledge base. The matching function characterizes the similarity between the semantic features of the target and the semantic centers of the known categories. The specific expression of the matching function is: , in, For the matching function, Let be the semantic features of the target, and ‖·‖ be the Euclidean norm.

[0025] Extract the category label of the known category corresponding to the maximum value of the matching function, and output it as the matched known category.

[0026] Furthermore, the specific steps of S2 are as follows:

[0027] A pulsed laser beam from a pulsed laser is directed onto the target, causing the target to reflect a pulse echo. The pulse echo is received using a single-photon detector. A time-to-digital converter is used to convert the pulse echo into a time histogram.

[0028] The present invention also provides a single-photon target recognition device based on a semantic knowledge base. The device uses the aforementioned single-photon target recognition method to identify targets. The device includes:

[0029] A pulsed laser is used to output pulsed laser light to irradiate the target.

[0030] A single-photon detector is used to receive the pulse echo reflected by the target.

[0031] A time-to-digital converter is used to convert the pulse echo into a time histogram.

[0032] The signal processor includes an encoder, a detection matcher, and a semantic knowledge base. The encoder is used to extract semantic features from the time histogram. The semantic knowledge base is used to store and retrieve category semantic centers, where each category semantic center is the average value of the semantic features within the same category. The detection matcher is used to perform unknown detection on the semantic features against the semantic knowledge base to determine whether the semantic features are known relative to the semantic knowledge base.

[0033] The detection matcher is further configured to perform semantic matching between the semantic features and the semantic knowledge base, and output the category label of the matched known category as the recognition result.

[0034] The detection matcher is also used to add the category semantic center of the new category to the semantic knowledge base to update the semantic knowledge base.

[0035] Preferably, the encoder is obtained by training a neural network model, the neural network model adopts a probabilistic step variational autoencoder structure, and the neural network model includes the encoder, decoder and loss function.

[0036] During training, the encoder maps the temporal histogram to the latent space to obtain a mean vector and a standard deviation vector, which are then output to the decoder. The encoder takes the mean vector as the semantic feature. The encoder also performs one-hot encoding on the category label, mapping it from the fully connected layer of the encoder to the latent space to obtain a predicted semantic center, which is then output to the decoder.

[0037] The decoder is used to sample the latent representation distribution based on the mean vector and the standard deviation vector, and to calculate the reconstructed histogram output based on the latent representation distribution. The decoder is also used to calculate the similarity between the semantic features and the predicted semantic centers, and to take the category label corresponding to the maximum similarity as the predicted category label output.

[0038] The predicted semantic center corresponds one-to-one with the category label. The latent representation distribution conforms to a Gaussian distribution centered at the mean vector and with the standard deviation vector as the radius. The reconstructed histogram is the reconstructed time histogram.

[0039] The technical solution provided by this invention has the following beneficial effects:

[0040] (1) By introducing unknown detection, this invention can clearly distinguish between known and unknown targets, avoiding the problem of forcibly misclassifying unknown targets as known categories. Unknown detection also triggers online incremental learning of the semantic knowledge base, enabling this method to dynamically transform unknown new targets into known ones without retraining the entire neural network model. This achieves continuous updating and adaptation of the semantic knowledge base in a real and open environment, solving the fundamental problem of the difficulty in updating traditional deep learning models after deployment.

[0041] (2) The present invention matches the inherent uncertainty of the signal with probabilistic modeling, which fundamentally enhances the robustness of the latent representation distribution. It adopts a probabilistic step variational autoencoder structure to probabilistically model the single-photon time histogram with high randomness and sparsity, and learns the latent representation distribution with semantic features as the center and the standard deviation of feature uncertainty as the radius. The noise and information incompleteness of the signal itself are integrated into the latent representation distribution, so that the latent representation distribution can remain stable even when the input signal is weak and the noise is severe. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the single-photon target recognition method according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram illustrating the process of constructing a semantic knowledge base according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the process for obtaining a time histogram according to an embodiment of the present invention;

[0045] Figure 4 This is a flowchart illustrating the semantic knowledge base self-updating mechanism according to an embodiment of the present invention;

[0046] Figure 5 This is a comparison diagram of the recognition effect of the embodiment of the present invention and the conventional method; Figure 5 (a) is a comparison chart of the recognition performance of the embodiment of the present invention and the conventional method at different collection times; Figure 5 (b) is a comparison chart of the recognition performance of the embodiments of the present invention and conventional methods under different signal-to-noise ratios;

[0047] Figure 6 This is a comparison chart of the recognition effect before and after the semantic knowledge base update in an embodiment of the present invention; Figure 6 (a) is a comparison chart of the recognition effect before and after the semantic knowledge base is updated in a real-world scenario; Figure 6 (b) is a comparison chart of the recognition effect before and after the semantic knowledge base is updated in a simulation scenario;

[0048] Figure 7 This is a comparison chart of the recognition effects of the semantic knowledge base before and after updating under different conditions according to an embodiment of the present invention; Figure 7 (a) is a comparison of the recognition performance before and after the semantic knowledge base update under different collection times and different signal-to-noise ratios; Figure 7 (b) is a comparison chart of the recognition effect before and after the semantic knowledge base is updated in real data and simulated data;

[0049] Figure 8 This is a schematic diagram of the structure of a single-photon target recognition device according to an embodiment of the present invention.

[0050] The labels in the diagram represent:

[0051] 1. Pulsed laser; 2. Pulsed laser; 3. Target; 4. Noise; 5. Single-photon detector; 6. Time histogram. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments of this invention will be described in further detail below with reference to examples. It should be understood that the embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.

[0053] Example 1

[0054] Please see the appendix Figure 1 , Figure 1 This is a schematic flowchart of a single-photon target recognition method according to an embodiment of the present invention. As shown in the figure, the single-photon target recognition method includes the following steps:

[0055] S1: Train the neural network model and construct a semantic knowledge base. The neural network model adopts a probabilistic ladder variational autoencoder structure to obtain semantic knowledge from the time histogram. Figure 6 Semantic features are extracted. Based on the extracted semantic features, to effectively distinguish different categories, the idea of ​​Conditional Gaussian Distribution Learning (CGDL) is adopted. The semantic features of different categories are modeled as different multivariate Gaussian distributions in the latent space, so that each category has a unique Gaussian distribution, which enhances the inter-class separability. For semantic features of the same category, the average value is taken as the category semantic center to construct a semantic knowledge base (SBK), thereby enhancing the intra-class compactness.

[0056] S2: Obtain the time histogram of target 3 Figure 6 .

[0057] S3: Calculate the time histogram of target 3 Figure 6 Input the neural network model to extract the semantic features of the target 3.

[0058] S4: Perform unknown detection on the semantic features of target 3 to determine whether target 3 is known relative to the semantic knowledge base. If yes, proceed to step S5. If no, proceed to step S6.

[0059] S5: Perform semantic matching on the semantic features of the target 3 relative to the semantic knowledge base, and output the matched known category as the recognition result.

[0060] S6: Take the average of the semantic features of the target 3 as the semantic center of the new category, add it to the semantic knowledge base to update the semantic knowledge base, and output the new category as the recognition result.

[0061] For example, the neural network model includes an encoder and a decoder. The encoder converts the received time histogram of target 3... Figure 6 Encoded as a pair of latent vectors in the latent space: mean vector and standard deviation vector The mean vector and standard deviation vector are used as parameters of the multivariate Gaussian distribution. In this embodiment of the invention, known category labels are manually set and input into the neural network model. The encoder performs one-hot encoding of the category labels, and the fully connected layer of the encoder maps them to the latent space, thereby obtaining the known categories corresponding one-to-one with the category labels. Predictive semantic center The predicted semantic center is then output to the decoder. In one embodiment of the present invention, during training, the decoder uses a reparameterization technique to sample the latent representation distribution in the latent space with the mean vector as the center and the standard deviation vector as the radius. ,in The decoder computes a reconstructed histogram based on the latent representation distribution, ensuring that the learned semantic features have a meaningful information structure. This reconstructed histogram is the temporal histogram reconstructed by the decoder. Figure 6 The decoder is also used to calculate the similarity between the semantic features of target 3 and the predicted semantic center, and to take the category label corresponding to the maximum similarity as the predicted category label output.

[0062] For example, the neural network model further includes a loss function. The loss function of the neural network model includes: reconstruction loss. divergence loss and classification loss Reconstruction loss Used to calculate the time histogram using the L2 norm. Figure 6 The difference between the reconstructed temporal histogram and the one used to ensure that the latent representation retains the temporal histogram of objective 3 with respect to the input is as follows: Figure 6 Sufficient information is obtained. A Gaussian distribution centered on the predicted semantic center and with the standard deviation vector as its radius is used as the reference distribution. Divergence loss. This is used to characterize the difference between the latent representation distribution and the reference distribution, making Compared with the reference distribution Alignment enhances intra-class compactness and inter-class separation in the latent space, thereby improving generalization ability. Classification loss. This is used to characterize the difference between the output predicted class label and the input class label. In one embodiment of the invention, the classification loss is specifically the softmax cross-entropy loss between the predicted class label and the class label. The loss function of the neural network model is specifically expressed as: , in, The scheduling parameter is linearly increased from 0 to 1 during training. These are fixed weights used to balance the classification loss.

[0063] For example, in one embodiment of the present invention, step S1 specifically includes the following steps:

[0064] S11: Collect a training sample set, which includes several sets of sample data, each set of sample data including the time histogram. Figure 6 And the category label. S12: Input the training sample set into the aforementioned neural network model, minimize the loss function through iterative evolution, so that the output latent representation distribution is as aligned as possible with the reference distribution, and the output reconstructed histogram is as close as possible to the input time histogram. Figure 6 The predicted category label is output to match the input category label. S13: After training, the category semantic center is output to the semantic knowledge base to construct the semantic knowledge base.

[0065] Please see the appendix Figure 2 , Figure 2 This is a schematic diagram illustrating the process of constructing a semantic knowledge base according to an embodiment of the present invention. Exemplarily, the steps of constructing a semantic knowledge base using a neural network model include: [The following is a partial translation of the original text, which is incomplete and requires further context.] Figure 6 Input a neural network model to extract semantic features corresponding to known categories. In this embodiment of the invention, for each known category, a set of semantic features belonging to that known category is collected, and the mean vector of the semantic features in the set is calculated as the category semantic center of the known category. The specific formula for calculating the category semantic center is as follows: , in, Let m be the semantic center of the m-th known category in the semantic knowledge base. Let m be the set of semantic features in the m-th known category. For the m-th known category, the first... A vector of mean values.

[0066] In this embodiment of the invention, to capture intra-class differences, it is also necessary to calculate the variance vector of each known class: Ultimately, the semantic knowledge base is defined as the set of category semantic centers for all known categories: The semantic knowledge base serves as a reference for semantic matching in the target recognition stage, supporting efficient and robust recognition.

[0067] For example, please refer to the appendix. Figure 3 , Figure 3 This is a schematic flowchart illustrating the process of obtaining a time histogram according to an embodiment of the present invention. Obtaining the time histogram... Figure 6 The steps include: at the source end, the target 3 is irradiated by a pulsed laser 2 output from a pulsed laser 1, the expression for which is: , in, For pulsed laser 2, The width of pulse laser 2, Let be the peak intensity of pulsed laser 2. The time characteristics of target 3 are denoted as . To describe the influence of target 3's geometry and surface reflectivity on the pulse echo, the specific representation of target 3's temporal characteristics is obtained as follows: , in, For Dirac functions, At the speed of light, Let f be the distance between target 3 and the detector. Mapping f transforms the information of target 3 from the spatial domain to the time domain to form a transmission signal. ,Right now The transmitted signal is the time histogram sent from the source end, and its specific form is as follows: , in, For convolution. When When it is much less than 1, It can be approximated as ,Right now At this time, the transmitted signal .

[0068] For example, various noises in the detection system can cause the transmitted signal to... Distortion occurs during channel transmission, resulting in the received signal. In this embodiment of the invention, the received signal This refers to obtaining the time histogram of target 3. Figure 6 There are two main types of noise: one is the detector's time jitter, which is the uncertainty in the detection system's time, denoted as... Another type is background noise, which includes background noise caused by ambient light. And the dark count caused by the detector .

[0069] For example, the time jitter of the detector is modeled as a Gaussian distribution: , in, Let be the standard deviation of the detector's time jitter. Therefore, the probability distribution of the received signal is: .

[0070] When the pulse echo signal is extremely weak, the photon count rate of pulsed laser 2 at time t is: , in, For the detector's detection efficiency. This represents the average number of photons reaching the detector from the pulse echo. At that time, photon counting can be simplified to: .

[0071] For example, the specific steps of S4 include:

[0072] The m-th known category is obtained from the neural network model and modeled as a multivariate Gaussian distribution in the latent space. and diagonal covariance matrix ,in, Let be the semantic center of the m-th known category. For the semantic features of target 3, its likelihood relative to the m-th known category is expressed as: , In the formula, Let be the likelihood function. The semantic features of the target. The category semantic center of the m-th known category The first component, The category semantic center of the m-th known category The d-th component, Semantic features for target 3 The first component, Semantic features for target 3 The d-th component, For the m-th known category, a multivariate Gaussian distribution is given. Let be the vector of integral variables in the latent space. The vector of integral variables in the latent space The first component, The vector of integral variables in the latent space The d-th component.

[0073] If the semantic features of target 3 do not exceed the likelihood threshold relative to all known categories, then target 3 is determined to be unknown relative to the semantic knowledge base. Based on the determination result, semantic matching or semantic knowledge base update is performed to complete target recognition.

[0074] In one embodiment of the present invention, the single-photon target recognition process is equivalent to a semantic communication model based on a semantic knowledge base. This allows the system to utilize a semantic matching method based on the semantic knowledge base to transform target recognition into a semantic matching process. This not only improves the robustness of the present invention's technical solution under low signal-to-noise ratio and short acquisition time conditions, but also enables the semantic knowledge base to dynamically adapt to new categories, enhancing its scalability and adaptability. Please refer to the appendix. Figure 4 , Figure 4This is a flowchart illustrating the semantic knowledge base self-updating mechanism of this invention. For example, if the maximum value in the likelihood function is greater than the likelihood threshold, then target 3 is judged to be known relative to the semantic knowledge base in the unknown detection, and semantic matching is performed to determine the recognition result. The specific steps of semantic matching include: calculating the matching function of the semantic features of target 3 relative to the known categories, whereby the matching function characterizes the similarity between the semantic features of target 3 and the category semantic centers of the known categories. The specific expression of the matching function is: , in, For the matching function, Let ||·|| represent the semantic features of objective 3, and ||·|| represent the Euclidean norm. The category label of the known category corresponding to the maximum value of the matching function is extracted as the known category output for the matching. In one embodiment of this invention, the category label of the corresponding known category is called the predicted category label, and the formula for extracting the predicted category label is as follows: , In the formula, To predict category labels, A set of semantic centers of several categories. The time histogram of objective 3 Figure 6 semantic features of target 3 The mapping function.

[0075] For example, if target 3 is determined to be unknown relative to the semantic knowledge base through unknown detection, then the semantic features of target 3 are used to update the semantic knowledge base. In one embodiment of the present invention, the step of updating the semantic knowledge base includes: calculating the average of the semantic features of all targets 3 belonging to the same unknown category, and using it as the category semantic center of the new category. The specific formula for calculating the category semantic center of the new category is as follows: , in, As the category semantic center of the new category, The number of semantic features of target 3 belonging to the same unknown category. The semantic features of target 3 are defined for the unknown category. The semantic features of the new category are added to the semantic knowledge base, thus enabling the semantic knowledge base to function from its initial state. arrive The self-updating mechanism enables the semantic knowledge base to be continuously updated without retraining, greatly enhancing the system's adaptability to new target categories in open set environments.

[0076] In one or other embodiments of the present invention, the single-photon target recognition method based on a semantic knowledge base can be applied to different scenarios. For example, in intelligent security systems, single-photon detection systems are deployed in high-security areas such as border defense, airports, and ports. The single-photon target recognition method of the present invention enables rapid target recognition, maintaining high accuracy, especially at night or under complex weather conditions. For unmanned driving perception systems, the single-photon recognition system or method provided by the embodiments of the present invention can promptly identify or update the category of sudden unknown obstacles or environmental targets, such as fallen traffic signs or animals, to assist in decision-making. In the field of aerospace remote sensing and military monitoring applications, for identifying flying targets in unknown battlefields or space scenarios, the single-photon recognition system or method provided by the embodiments of the present invention can achieve remote, low-power target perception and recognition without relying on high-definition images. In resource-constrained applications, such as micro-robots and sensor network nodes, where traditional camera systems cannot be deployed due to power consumption and size limitations, the single-photon target recognition system or method of the embodiments of the present invention can achieve lightweight target perception. In industrial inspection applications, such as high-speed production line inspection, the single-photon recognition system or method provided in this invention can quickly adapt to the first-time occurrence of defect sample categories, solving the technical difficulties of slow or incorrect recognition by traditional models.

[0077] In one or other embodiments of the present invention, the single-photon target recognition method based on a semantic knowledge base can solve the problems of existing data-driven single-pixel single-photon detection (DPS). 2 SP 2 D) Target recognition technology suffers from significant performance degradation under low signal-to-noise ratio (SNR) and short acquisition time conditions, and it cannot identify or adapt to unknown targets in open-set recognition scenarios. Data-driven single-pixel single-photon detection offers a promising solution for long-range, low-size, low-cost, and low-power imaging-free target recognition. However, when noise is strong or acquisition time is short, the number of received photons decreases, and the time histogram... Figure 6 With a sharp reduction in effective information, conventional analysis and recognition based on image or histogram features using deep learning models (such as ResNet-18 and U-Net) are prone to failure, resulting in a significant drop in recognition accuracy.

[0078] In a closed-set recognition scenario, the recognition performance of the single-photon target recognition method in one embodiment of the present invention was compared with that of existing technologies under different signal-to-noise ratios and acquisition time conditions. Please refer to the appendix. Figure 5 , Figure 5This is a comparison chart of the recognition performance of the embodiments of the present invention with conventional methods. First, without considering unknown categories, the single-photon target recognition method of the embodiments of the present invention is compared with two conventional deep learning model recognition methods—ResNet-18 and U-Net. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio. Both ResNet-18 and U-Net are trained using the Softmax cross-entropy loss function. In the methods using ResNet-18 and U-Net, the received signal x is directly input into the neural network model, and the predicted category label is output. For the single-photon target recognition method based on a semantic knowledge base proposed in this invention, the validation set is used to construct the semantic knowledge base. Figure 5 (a) is a comparison of the recognition performance of the embodiments of the present invention with conventional methods under different acquisition times. The method of the embodiments of the present invention is significantly better than the existing recognition methods ResNet-18 and U-Net. At shorter acquisition times (e.g., less than 10ms), the method of the embodiments of the present invention shows a significant advantage and higher accuracy. When the acquisition time increases (e.g., more than 50ms), the accuracy of each method tends to be consistent, but the method of the embodiments of the present invention still has a slight advantage. Figure 5 (b) is a comparison of the recognition performance of the embodiments of the present invention and conventional methods under different signal-to-noise ratios. The method of the embodiments of the present invention exhibits stronger robustness to noise 4 and outperforms ResNet-18 under various signal-to-noise ratio conditions. When the signal-to-noise ratio is -15dB, the recognition accuracy of the method of the embodiments of the present invention is more than 25% higher than that of ResNet-18 and more than 50% higher than that of U-Net.

[0079] In conventional methods, when a deep learning model encounters a new category in the test set that is not covered by the training set, it often forcibly classifies it into a known category, failing to correctly identify the unknown target.3 This misidentification poses a serious threat to security defense and autonomous decision-making systems. (See appendix) Figure 6 , Figure 6 This is a comparison chart of the recognition performance before and after the semantic knowledge base update in an embodiment of the present invention. To verify the self-updating function of the single-photon target recognition method based on the semantic knowledge base in handling unknown category scenarios, the performance before and after the semantic knowledge base update was compared. Figure 6 (a) is a comparison chart of recognition performance before and after the semantic knowledge base update in a real-world scenario. Figure 6(b) is a comparison of the recognition performance before and after the semantic knowledge base update in a simulated scenario. In this embodiment of the invention, two experimental scenarios were designed based on real and simulated datasets: in the real scenario, 30 categories were considered known categories and 1 was an unknown category; in the simulated scenario, 40 categories were considered known categories and 1 was an unknown category. The unknown category in the real dataset is "adult back". Before the update, the semantic knowledge base misclassified this unknown category as one of the known categories. After the update, the semantic knowledge base can accurately identify the "adult back" category while maintaining a high recognition rate for known categories. The harmonic mean of precision and recall was calculated to obtain the F1 score, and it was found that the F1 score of the updated semantic knowledge base for target recognition was significantly improved, indicating a comprehensive improvement in classification performance. The results based on the synthetic dataset also showed similar performance improvement. In addition to the F1 scores of each category, the macro-average F1 score (the arithmetic mean of the F1 scores of all categories) before and after the update was also evaluated. The results show that the macro average F1 score in the real dataset increased from 0.8474 to 0.9319, and in the simulation dataset it increased from 0.9561 to 0.9676. This further verifies the effectiveness of the self-updating mechanism of the semantic knowledge base in improving the overall recognition performance. In particular, it enables robust open set recognition without retraining the neural network model when the unknown category is included.

[0080] Please see the appendix Figure 7 , Figure 7 This is a comparison chart of the recognition performance of the semantic knowledge base before and after updating under different conditions according to an embodiment of the present invention. In one embodiment of the present invention, the performance of the semantic knowledge base before and after updating was evaluated under different signal-to-noise ratios and acquisition times. Figure 7 (a) shows a comparison of the recognition performance of the semantic knowledge base before and after the update under different acquisition times and signal-to-noise ratios. Under different acquisition times, the updated semantic knowledge base shows a significant improvement in recognition accuracy, indicating its ability to adapt to new categories in real time. When the signal-to-noise ratio is higher than -13dB, the updated semantic knowledge base consistently outperforms the unupdated version. As the signal-to-noise ratio decreases, the performance gap between the two narrows, but the updated semantic knowledge base still maintains a slight advantage. This demonstrates that the self-updating mechanism can maintain good recognition capabilities in high-noise environments, making it particularly important in real-world applications.

[0081] In one embodiment of the present invention, as the number of unknown categories increases, the performance of the semantic knowledge base self-updating on real and synthetic data is further tested. Figure 7(b) shows a comparison of the recognition performance before and after the semantic knowledge base update in real and simulated data. The recognition accuracy of the unupdated semantic knowledge base drops rapidly: when the number of unknown classes increases to 9, the accuracy of the experimental data is below 70%, while that of the simulated data is close to 80%. The updated semantic knowledge base, however, maintains an accuracy of over 90%, demonstrating good scalability. Although the accuracy of the semantic knowledge base method fluctuates slightly as the number of unknown classes increases, it consistently remains above 90%.

[0082] Example 2

[0083] Embodiment 2 of this invention proposes a single-photon target recognition device based on a semantic knowledge base, which uses the aforementioned single-photon target recognition method to identify target 3. Please refer to the appendix. Figure 5 , Figure 5 This is a schematic diagram of the structure of a single-photon target recognition device according to an embodiment of the present invention. The device includes: a pulsed laser 1, a single-photon detector 5, a time-to-digital converter, and a signal processor. The signal processor further includes an encoder, a detection matcher, and a semantic knowledge base.

[0084] Exemplarily, a pulsed laser 1 is used to output a pulsed laser 2 to illuminate the target 3. A single-photon detector 5 is used to receive the pulse echo reflected by the target 3. In one or other embodiments of the invention, the single-photon detector 5 is specifically a single-photon avalanche diode. A time-to-digital converter is used to convert the pulse echo into a time histogram. Figure 6 The encoder is used to extract the time histogram. Figure 6 Semantic features are extracted. The semantic knowledge base is used to store and retrieve category semantic centers, where each category semantic center is the average value of the semantic features within the same category. The detection matcher is used to perform unknown detection on the semantic features against the semantic knowledge base to determine whether the semantic features are known relative to the semantic knowledge base. The detection matcher is also used to perform semantic matching on the semantic features against the semantic knowledge base, outputting the category label of the matched known category as the recognition result. The detection matcher is also used to add the category semantic center of the new category to the semantic knowledge base to update the semantic knowledge base.

[0085] In this embodiment of the invention, the neural network model employs a probabilistic step variational autoencoder structure, and the neural network model includes the encoder, decoder, and loss function. During training, the encoder is used to convert the time histogram... Figure 6The semantic features are encoded into the latent space to obtain a mean vector and a standard deviation vector, which are then output to the decoder. The encoder takes the mean vector as the semantic feature. The encoder is also used to perform one-hot encoding of the category label, which is mapped to the latent space by a fully connected layer to obtain a predicted semantic center, which is then output to the decoder. The decoder samples the latent representation distribution with the mean vector as the center and the standard deviation vector as the radius, and calculates a reconstruction histogram based on the latent representation distribution. The decoder is also used to calculate the similarity between the semantic feature and the predicted semantic center, and takes the category label corresponding to the maximum similarity as the predicted category label. The predicted semantic center corresponds one-to-one with the category label. The reconstruction histogram is the reconstructed time histogram. Figure 6 .

[0086] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for single-photon target recognition based on a semantic knowledge base, characterized in that, The method comprises the following steps: S1: training a neural network model to build a semantic knowledge base; the neural network model adopts a probabilistic ladder variational autoencoder structure, is used for extracting semantic features from a time histogram, and models the semantic features of different categories as a multivariate Gaussian distribution, takes the average value of the semantic features of the same category as a category semantic center, and stores in the semantic knowledge base; S2: obtaining a time histogram of a target; S3: inputting the time histogram of the target into the neural network model to extract semantic features of the target; S4: performing unknown detection on the semantic features of the target, and judging whether the target is known relative to the semantic knowledge base, if yes, jumping to step S5; if no, jump to step S6; S5: performing semantic matching on the semantic features of the target relative to the semantic knowledge base, and outputting a matched known category as a recognition result; S6: taking the average value of the semantic features of the target as the category semantic center of a new category, adding the new category to the semantic knowledge base to update the semantic knowledge base, and outputting the new category as the recognition result.

2. The single-photon object recognition method of claim 1, wherein, The neural network model comprises an encoder and a decoder; The encoder is used for encoding the time histogram to a latent space to obtain a mean vector and a standard deviation vector, and outputting to the decoder; the encoder takes the mean vector as the semantic feature; The encoder is also used for one-hot encoding of a category label, mapping to the latent space by a fully connected layer of the encoder to obtain a predicted semantic center, and outputting to the decoder; The decoder is used for sampling a latent representation distribution according to the mean vector and the standard deviation vector, and calculating a reconstructed histogram output according to the latent representation distribution; the decoder is also used for calculating the similarity of the semantic features and the predicted semantic center, and taking the category label corresponding to the maximum similarity as a predicted category label output; Wherein, the predicted semantic center and the category label correspond one by one; the latent representation distribution conforms to a Gaussian distribution with the mean vector as the center and the standard deviation vector as the radius; the reconstructed histogram is the reconstructed time histogram.

3. The single-photon object recognition method of claim 2, wherein, The neural network model further comprises a loss function, The loss function comprises a divergence loss, a reconstruction loss and a classification loss; the reconstruction loss is used to represent the difference between the time histogram and the reconstructed histogram; the divergence loss is used to represent the difference between the latent representation distribution and a reference distribution; The classification loss is used to represent the difference between the predicted category label output and the category label input; Wherein, the reference distribution conforms to a Gaussian distribution with the predicted semantic center as the center and the standard deviation vector as the radius.

4. The single-photon object recognition method of claim 3, wherein, The specific steps of S1 are: S11: a training sample set is collected, the training sample set includes a plurality of groups of sample data, each group of sample data includes the time histogram and the category label; S12: the training sample set is input into the neural network model, the loss function is minimized in an iterative evolution manner, the output latent representation distribution is aligned with the reference distribution as much as possible, the output reconstructed histogram is as close to the input time histogram as possible, and the output predicted category label conforms to the input category label; S13: after the training is completed, the category semantic center is output to the semantic knowledge base, so as to construct the semantic knowledge base.

5. The single-photon target identification method of claim 1, wherein, The specific steps of S4 are: S41: the likelihood function of the semantic feature of the target with respect to each known category in the semantic knowledge base is calculated respectively, so as to represent the probability that the target belongs to the known category respectively, and the likelihood function is specifically represented as: ， In the formula, Let be the likelihood function. The semantic features of the target. The semantic center of the m-th known category is given. The category semantic center of the m-th known category The first component, The category semantic center of the m-th known category The d-th component, semantic features of the target The first component, semantic features of the target The d-th component, For the m-th known category, a multivariate Gaussian distribution is given. Let be the vector of integral variables in the latent space. The vector of integral variables in the latent space The first component, The vector of integral variables in the latent space The d-th component; S42: if the maximum value in each likelihood function is greater than a likelihood threshold, it is judged that the target is known with respect to the semantic knowledge base; otherwise, it is judged that the target is not known with respect to the semantic knowledge base.

6. The single-photon object recognition method of claim 5, wherein, The specific steps of S5 are: The matching function of the semantic feature of the target with respect to each known category in the semantic knowledge base is calculated respectively; the matching function is used to represent the similarity between the semantic feature of the target and the category semantic center of the known category; and the specific expression of the matching function is: ， wherein, is the matching function, is the semantic feature of the target, and ‖·‖ is the Euclidean norm. The category label of the known category corresponding to the maximum value of the matching function is extracted as the output of the matched known category.

7. The single-photon target identification method of claim 1, wherein, The specific steps of S2 are: Pulse laser output by a pulse laser is irradiated to the target, so that the target reflects to generate a pulse echo; A single-photon detector is used to receive the pulse echo; A time-to-digital converter is used to convert the pulse echo into the time histogram.

8. A single-photon target recognition device based on a semantic knowledge base, characterized in that, The device uses the single-photon target recognition method of claim 1 to recognize the target, and the device includes: A pulse laser is used to output pulse laser irradiation to the target; A single-photon detector is used to receive the pulse echo reflected by the target; A time-to-digital converter is used to convert the pulse echo into a time histogram; A signal processor includes an encoder, a detection matcher and a semantic knowledge base, the encoder is used to extract a semantic feature from the time histogram; the semantic knowledge base is used to store and call a category semantic center, the category semantic center is an average value of the semantic features of the same category; the detection matcher is used to perform unknown detection on the semantic feature and the semantic knowledge base, to determine whether the semantic feature is known with respect to the semantic knowledge base; The detection matcher is also used to perform semantic matching on the semantic feature and the semantic knowledge base, to output the category label of the matched known category as a recognition result; The detection matcher is also used to add the category semantic center of a new category to the semantic knowledge base, to update the semantic knowledge base.

9. The single-photon object recognition device of claim 8, wherein, The encoder is obtained from a trained neural network model, the neural network model adopts a probabilistic ladder variational autoencoder structure, and the neural network model comprises the encoder, a decoder and a loss function; During training, the encoder is configured to encode the time histogram into a latent space to obtain a mean vector and a standard deviation vector, and output to the decoder; the encoder takes the mean vector as the semantic feature; the encoder is further configured to one-hot encode a category label, map the one-hot encoded category label to the latent space by a fully connected layer of the encoder to obtain a predicted semantic center, and output to the decoder; The decoder is configured to sample a latent representation distribution according to the mean vector and the standard deviation vector, and calculate a reconstructed histogram output according to the latent representation distribution; the decoder is further configured to calculate a similarity between the semantic feature and the predicted semantic center, and take a category label corresponding to a maximum similarity as a predicted category label output; The predicted semantic center corresponds to the category label in a one-to-one manner; the latent representation distribution conforms to a Gaussian distribution with the mean vector as a center and the standard deviation vector as a radius; and the reconstructed histogram is a reconstructed time histogram.

Images