Infrared target recognition method and system based on multi-dimensional feature knowledge base and mixed Gaussian model

Abstract

The application discloses an infrared target recognition method and system based on a multi-dimensional feature knowledge base and a Gaussian mixture model, and the method comprises the following steps: a target detection network is used to extract a region of interest of an infrared target and output corresponding multi-channel feature maps and instance segmentation masks; statistical attribute features of the target are extracted according to the multi-channel feature maps, geometric attribute features of the target are extracted according to the instance segmentation masks, and a multi-dimensional feature knowledge base is constructed; feature data in the multi-dimensional feature knowledge base is input into the Gaussian mixture model, the covariance matrices of each Gaussian component in the Gaussian mixture model are set as spherical covariance matrices, an expectation maximization algorithm is used for iterative optimization, and a category recognition result of the infrared target is output. Compared with the prior art, the application improves the explainability of classification decision and the robustness in a complex scene by constructing a multi-dimensional feature knowledge base and using a spherical covariance constrained Gaussian mixture model for probability clustering.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and infrared image processing technology, and in particular to an infrared target recognition method and system based on a multidimensional feature knowledge base and a Gaussian mixture model. Background Technology

[0002] Computer vision is a visual cognitive computing system for complex scenes, built upon images acquired from peripheral devices as input and referencing environmental perception and semantic parsing techniques from biological vision. It utilizes biomimetic vision mechanisms to decouple target features and understand the scene. This technology is rapidly developing in both military and civilian fields, such as autonomous driving, infrared reconnaissance, and environmental detection. Compared to traditional visible light imaging technology, whose performance significantly decreases in insufficient lighting or with excessive background interference, infrared imaging technology can penetrate smoke by capturing the thermal radiation characteristics of targets and has strong backlight resistance, allowing it to operate normally at night or in adverse weather conditions.

[0003] However, infrared images are low-resolution and lack texture, making them easily obscured by background noise during detection, and difficult to extract features using traditional algorithms. Early infrared target detection methods mainly employed simple thresholding techniques, using preset grayscale values ​​to separate the target from the background, but these were limited to the grayscale difference between targets of a certain size and the background. Subsequently, morphological filtering and inter-frame differencing methods were proposed, utilizing the temporal stability of the background to suppress interference from dynamic backgrounds. In recent years, breakthroughs in traditional methods have mainly relied on background modeling, but the preprocessing process is extremely complex.

[0004] In recent years, deep learning methods, represented by the YOLO series, have transformed detection into a regression problem, achieving bounding box and class prediction through a single forward propagation, thus improving detection accuracy and speed. However, existing deep learning models still have the following limitations: First, the feature representation capability is weak: key details are easily lost when extracting low-resolution infrared features, and once the target is occluded in continuous frame detection, the (ID) is easily switched, which leads to the failure of motion attribute (such as speed, acceleration, trajectory) extraction, making it difficult to achieve stable tracking across frames. Second, the black box effect of deep models leads to extremely poor interpretability: end-to-end deep networks cannot explain why a target is identified as such, lacking decision-making transparency, which severely limits the traceability and credibility in high-safety scenarios such as autonomous driving. Third, it lacks anti-interference capability in complex scenarios: Under unknown or anti-interference conditions, the feature space distribution is extremely irregular, and traditional classifiers are prone to misclassifying noise as targets.

[0005] Furthermore, traditional target recognition and knowledge base technologies rely heavily on expert systems, symbolic logic, and structured rules to achieve recognition through predefined rules. However, these technologies are difficult to adapt to complex and ever-changing environments, and there is a lack of effective connection between existing deep learning technologies and structured knowledge base mechanisms.

[0006] A search revealed that Chinese Patent Publication No. CN119131088A discloses a method for detecting and tracking weak targets in infrared images based on a lightweight hypergraph network. This method extracts the geometric and semantic features of infrared images through Riemannian manifold learning, constructs a complete database of infrared targets using a semantic feature variational autoencoder, and then extracts high-order structural information through a hypergraph network to achieve target detection and tracking. However, this method still belongs to an end-to-end deep learning black box model, and the decision-making process of its hypergraph network lacks interpretability, making it difficult to trace the physical features of the classification results. Furthermore, in scenarios with small samples and strong interference, the high-order correlation modeling of the hypergraph network is prone to introducing background noise.

[0007] Therefore, how to improve the interpretability of classification decisions and the robustness in complex scenarios while ensuring the accuracy of infrared target recognition is a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an infrared target recognition method and system based on a multidimensional feature knowledge base and a Gaussian mixture model.

[0009] The objective of this invention can be achieved through the following technical solutions: According to a first aspect of the present invention, an infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model is provided, the method comprising: The process involves acquiring an infrared image to be processed, using a target detection network to extract the region of interest (ROI) of the target in the infrared image, and outputting the corresponding multi-channel feature map and instance segmentation mask. The statistical attribute features of the target are extracted based on the multi-channel feature map, and the geometric attribute features of the target are extracted based on the instance segmentation mask. The statistical attribute features and geometric attribute features are stored in a structured manner to construct a multi-dimensional feature knowledge base. The feature data in the multidimensional feature knowledge base is input into the Gaussian mixture model, and the covariance matrix of each Gaussian component in the Gaussian mixture model is set as a spherical covariance matrix. The expectation-maximization algorithm is used for iterative optimization, and the category recognition result of the infrared target is output.

[0010] As a preferred technical solution, the statistical attribute features are used to characterize the thermal radiation distribution characteristics of the target, including: gray mean, gray variance, gray energy, gray entropy, and histogram invariant moments.

[0011] As a preferred technical solution, the geometric attribute features are used to characterize the shape and contour invariance of the target, including: geometric invariant moments, aspect ratio, Fourier descriptor, roundness, and major axis deflection angle.

[0012] As a preferred technical solution, the spherical covariance matrix is ​​a diagonal matrix in which the variances of each dimension of the features within the same Gaussian component are equal, so that the clusters corresponding to each Gaussian distribution are spherically distributed in the feature space.

[0013] As a preferred technical solution, the update of the spherical covariance matrix is ​​as follows: based on the weighted responsibility value corresponding to each Gaussian component, the average of the squared Euclidean distances from the feature data points corresponding to each Gaussian component to the mean vector of that Gaussian component is calculated, and then normalized according to the feature dimension.

[0014] As a preferred technical solution, the output target category identification result includes: outputting the probability distribution of each target belonging to each Gaussian distribution category, and selecting the category corresponding to the highest probability as the final classification result.

[0015] As a preferred technical solution, the convergence condition of the expectation maximization algorithm is that the change in the log-likelihood function between two adjacent iterations is less than a preset threshold.

[0016] As a preferred technical solution, the multi-channel feature map is cropped from different feature levels through region of interest alignment operation, and then the mean value is taken along the channel dimension to generate a single-channel target feature map. Statistical attribute features are extracted based on the single-channel target feature map.

[0017] As a preferred technical solution, before inputting the feature data in the multidimensional feature knowledge base into the Gaussian mixture model, the feature data is Z-score standardized to eliminate the dimensional differences between statistical attribute features and geometric attribute features.

[0018] According to a second aspect of the present invention, a system for infrared target recognition based on a multidimensional feature knowledge base and a Gaussian mixture model is provided, comprising: The perception module is used to extract the region of interest of the target from the infrared image and obtain multi-channel feature maps and instance segmentation masks; The knowledge base construction module extracts statistical attribute features based on the multi-channel feature map, extracts geometric attribute features based on the instance segmentation mask of the region of interest, and concatenates the two to construct a multi-dimensional feature knowledge base. The recognition module is used to input feature data from the multidimensional feature knowledge base into a Gaussian mixture model for cluster recognition, wherein the covariance matrix of each Gaussian component of the Gaussian mixture model is constrained to a spherical covariance matrix, and the target's class probability is output.

[0019] Compared with the prior art, the present invention has the following advantages: 1. This invention constructs a multi-dimensional feature knowledge base that includes statistical and geometric attributes, enabling target classification decisions to be traced back to specific grayscale distributions, thermal radiation patterns, and contour geometric parameters. This breaks the black-box limitation of deep learning models and improves the interpretability of the classification decision process while ensuring the accuracy of infrared target recognition.

[0020] 2. This invention constrains the covariance matrix of each Gaussian component in the Gaussian mixture model to a spherical covariance matrix, limiting the features of each dimension within the same Gaussian component to be independent and have equal variances. This effectively suppresses noise correlation in high-dimensional features and avoids the overfitting problem of the full covariance matrix, thereby improving the anti-interference robustness and generalization ability of the recognition system in complex interference environments.

[0021] 3. While outputting detection results, the target detection network of this invention further extracts multi-scale feature maps from the intermediate layer. After precise cropping and channel averaging by RoI Align, a single-channel grayscale feature map is generated, realizing the effective connection between deep learning features and the backend physical knowledge base. On this basis, geometric features are extracted by combining instance segmentation masks, realizing the organic integration of deep learning perception and physical feature expression. While maintaining detection efficiency, it enhances the model's feature capture ability for infrared weak targets and blurred boundaries.

[0022] 4. This invention integrates the statistical properties of thermal radiation and the geometrically invariant properties of infrared targets for multi-dimensional feature characterization, effectively compensating for the weak feature expression caused by the low resolution and low texture of infrared images, and improving the feature discrimination of different types of infrared targets.

[0023] 5. This invention uses the expectation-maximization algorithm to iteratively optimize the Gaussian mixture model with spherical covariance constraints, and completes target category identification through soft assignment. The decision-making process is stable and reliable, does not rely on a large amount of manually labeled data, and is suitable for unsupervised clustering requirements in complex infrared scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] Example 1: like Figure 1As shown, this invention provides an infrared target recognition method based on a multi-dimensional feature knowledge base and a Gaussian mixture model, specifically including the following steps: The process involves acquiring an infrared image to be processed, extracting the region of interest of the infrared target using a target detection network, and outputting the corresponding multi-channel feature map and instance segmentation mask.

[0027] The statistical attribute features of the infrared target are extracted based on the multi-channel feature map, and the geometric attribute features of the target are extracted based on the instance segmentation mask. The statistical attribute features and geometric attribute features are then stored in a structured manner to construct a multi-dimensional feature knowledge base.

[0028] The feature data in the multidimensional feature knowledge base is input into the Gaussian mixture model, and the covariance matrix of each Gaussian component in the Gaussian mixture model is set as a spherical covariance matrix. The expectation-maximization algorithm is used for iterative optimization, and the category recognition result of the infrared target is output.

[0029] The method of the present invention will be described in detail below: Step S1: Infrared target perception and multi-channel feature map extraction.

[0030] The goal of this step is to detect the target from the raw infrared image and extract its multi-channel feature map, providing a data foundation for subsequent knowledge base construction.

[0031] Step S11: Data acquisition and preprocessing.

[0032] This embodiment uses an infrared dataset containing rich and complex scenes, such as the FLIR_ADAS_v2 dataset and the shortwave infrared ship target instance segmentation dataset. The original dataset annotation format is usually a JSON file in COCO format, which includes image resolution, bounding box coordinates, category labels and instance segmentation mask polygon coordinates.

[0033] First, load and parse the JSON annotation file to extract the image width. and height And the category mapping relationship for each target. For each target, its bounding box coordinates are... Convert to absolute coordinates For the polygon coordinates of the instance segmentation mask, their width and height are normalized. Convert the above absolute coordinates into the normalized center coordinates and width and height required for deep learning network training: , , , , The calculated Write the data into a TXT file and generate structured data containing mask contour normalized points in the same directory for subsequent network training.

[0034] Step S12: Construction and training of the object detection network.

[0035] This embodiment employs a pre-trained target detection network to extract the region of interest (ROI) of infrared targets. Preferably, the network uses the YOLOv11 model, with its backbone incorporating a C3K2 module and a SPPF (Spatial Pyramid Pooling - Fast) module. C3K2 reduces network volume and accelerates gradient backpropagation through residual edges; the SPPF module effectively increases the receptive field of infrared targets through multiple cascaded 5×5 max-pooling layers. Simultaneously, a C2PSA attention mechanism is integrated to guide the network to suppress infrared background noise and focus on areas of significant thermal radiation. The neck layer uses a PANet structure to achieve bidirectional feature fusion from top to bottom and bottom to top, concatenating deep high-level semantic information (for identifying target types) with shallow high-resolution information (for locating target edges). The head layer uses a decoupled structure, outputting a category prediction branch, a bounding box regression branch, and an instance segmentation mask branch.

[0036] During the training phase, the total loss function consists of three parts: , in, These are the weighting coefficients. For bounding box regression loss, For category classification loss, For instance segmentation mask loss, The CIoU (Complete Intersectionover Union) loss is employed, which considers not only the overlapping area but also the relative proportions of the center point distance and aspect ratio, thus improving the positioning accuracy under infrared blurred boundaries. After approximately 400 epochs of training, the optimal weight file best.pt is saved when the loss function converges.

[0037] The preprocessed infrared image is input into the network, which outputs the predicted category, confidence score, absolute coordinates of the bounding box, and instance segmentation mask for each target. The instance segmentation mask is a binary image, with pixel values ​​set to 255 within the target region and 0 within the background region.

[0038] Step S13: Region of Interest Alignment and Feature Map Generation.

[0039] For each detected target, its horizontal bounding box coordinates are extracted, a virtual batch_id is added, and it is converted into a tensor format of [N,5]. This tensor and the multi-scale feature maps of the intermediate layers of the network are then input into the RoI Align operator. RoIAlign uses a bilinear interpolation algorithm to avoid pixel misalignment caused by two quantizations. The output size is set to... In this embodiment, 64×64 pixels are used. For the corresponding points on the feature map... Its pixel value is calculated by bilinear interpolation from the four real pixels surrounding it.

[0040] Multiple feature blocks are obtained by cropping from feature maps at different levels, and their tensor format is as follows: ,in K This represents the sum of the number of channels in the feature maps at each level. These feature blocks are concatenated along the channel dimension to obtain a multi-channel fused feature tensor. Then, the mean is calculated along the channel dimension to reduce the dimensionality and generate a single-channel grayscale target feature map in the format [1,1,64,64]. This map is then converted into a matrix array (e.g., in .npy format) and saved as the data base for subsequent physical feature extraction. This single-channel grayscale target feature map is the basic data used to extract statistical attribute features.

[0041] Step S2: Construction of a multi-dimensional target feature knowledge base This step is based on the single-channel grayscale target feature map (64×64) and instance segmentation mask obtained in step S1. The statistical attribute features and geometric attribute features of the target are extracted respectively, and all features are stored in a structured manner to build a multi-dimensional feature knowledge base.

[0042] Step S21: Extract statistical attribute features.

[0043] For a single-channel feature map, statistically analyze its gray-level histogram distribution, let... This represents the total number of pixels in the 64×64 feature map (i.e., 4096). grayscale The number of pixels that appear pixel grayscale value Given the normalized histogram probability, calculate the following statistical characteristics:

[0044] Gray mean Used to reflect the overall brightness level within the target area. ; in, For the specific pixel grayscale value in the feature map, This represents the frequency of the corresponding pixel grayscale value appearing within the target area. This represents the total number of pixels within the target area. Gray variance Used to measure the degree of deviation of pixel values ​​from the mean within a target area and image contrast. ; Grayscale energy Used to reflect the coarseness and uniformity of infrared texture. .

[0045] Gray entropy Used to measure the randomness or complexity of infrared target texture. ; Histogram higher-order invariant moments: Statistical moments calculated based on the gray-level histogram distribution of infrared targets, including first-order moments. Second-order central moments Third-order moment standardized skewness and fourth-order moment normalized kurtosis This is used to comprehensively describe the radiation distribution pattern of an infrared target, where skewness is: kurtosis: .

[0046] The above statistical characteristics are used together to characterize the thermal radiation distribution of infrared targets.

[0047] Step S22: Geometric attribute feature extraction.

[0048] For the binary mask image, the cv2.findContours function is called to extract the outer contour set. Based on the contour coordinates, the following calculations are performed: Seven geometrically invariant moments (Hu moments): First, calculate the p+q order two-dimensional spatial moments of the image: , Where p and q are the orders of the two-dimensional space moments. This represents the pixel value of the image at coordinates (x, y). Calculate the centroid coordinates of the target: ; Calculate the central moment: ; Calculate the normalized central moments: ; Finally, seven Hu moments were constructed that not only possess translation and rotation invariance but also scale invariance: , , , , , , , Fourier descriptor: using contour coordinates as points in the complex plane , where n is the index of the discrete points of the target contour; after Fast Fourier Transform (FFT), the amplitude of its low-frequency coefficients is truncated to remove high-frequency noise; Basic morphological features: Calculating the aspect ratio of the bounding box Where W is the width of the minimum bounding rectangle of the target, and H is the height of the minimum bounding rectangle of the target; roundness ,in, The area within the outline, The perimeter; and the major axis deflection angle. The tilt angle of the principal axis of the target shape relative to the coordinate axis is represented by the following formula: ,in For the covariance moment, and It is a second moment; The aforementioned geometric features are used together to characterize the shape profile invariance of infrared targets.

[0049] Step S23: Structured storage of the knowledge base.

[0050] The statistical attribute features (grayscale mean, variance, energy, entropy, skewness, kurtosis) and geometric attribute features (7 Hu moments, Fourier descriptors, aspect ratio, roundness, major axis skewness) extracted from each target are concatenated into a D-dimensional feature vector. Simultaneously, the target ID, detection category (the category directly output by the detection network, optional), and storage path of the instance segmentation mask are associated with the target. All feature vectors are organized by row to form a structured knowledge base, which can be stored in a database or in-memory matrix.

[0051] Step S3: Clustering identification based on spherical covariance matrix and Gaussian mixture model.

[0052] This step inputs the feature data from the multidimensional feature knowledge base constructed in step S2 into the Gaussian mixture model, applies a spherical covariance constraint to the model, and performs iterative optimization through the expectation-maximization (EM) algorithm to finally output the category recognition result of the infrared target.

[0053] Step S31: Data standardization.

[0054] To eliminate the dimensional differences between statistical and geometric features, Z-score standardization is applied to all feature data in the knowledge base: , in, For the first The first sample Dimensional features, and These are the mean and standard deviation of this dimension across the entire dataset, respectively.

[0055] Step S32: Initialize the Gaussian mixture model.

[0056] GMM assumes that all feature data are generated by It is composed of a linear combination of multidimensional Gaussian distributions, and its probability density function is: , The K-Means algorithm is used to perform preliminary clustering on the standardized feature data. The centroids, covariances, and sample proportions of the initial clusters are used to initialize the parameters of the Gaussian mixture model: the mean vector. Covariance matrix (Initialized as a full covariance matrix, then forcibly converted to a spherical shape) and mixed weights .

[0057] Step S33: Spherical covariance constraint.

[0058] This invention forces the covariance matrix of each Gaussian component to be a spherical covariance matrix, that is... Where I is the identity matrix, The constraint is a scalar variance. It assumes that the features in each dimension are independent of each other and have the same variance, so that each cluster is spherically distributed in the feature space. This constraint can effectively eliminate redundant correlations in high-dimensional features, while avoiding overfitting problems caused by the full covariance matrix. It is particularly suitable for scenarios where the natural correlation between the statistical features and geometric features of infrared targets is low.

[0059] This invention, through extensive experimental comparisons, has found that: Using tied (shared) covariance: all clusters are forced to share a single covariance shape, resulting in a complete loss of inter-cluster specificity; the adjusted RAND index ARI is only 0.013. Using full (arbitrary) covariance: the computational cost increases quadratically, and it is prone to overfitting under high-dimensional features. Using spherical covariance: it assumes that the features in each dimension are independent and isotropic, perfectly matching the physical prior that the extracted geometric and statistical features have low correlation. Its test set ARI reaches 0.083, and the normalized mutual information (NMI) exceeds 0.156. When facing infrared noise interference, the cluster boundaries are the smoothest and clearest, with the lowest misclassification rate.

[0060] Step S34: Iterative optimization of the EM algorithm.

[0061] The expectation-maximization algorithm is used to estimate the parameters of a spherical Gaussian mixture model. Let the feature data point set be... The posterior probability (responsibility value) of each point belonging to the k-th Gaussian component is: The iterative process is as follows: E-step (expectation step): Calculates the responsibility value for each data point generated by the k-th Gaussian distribution. , Where D is the total dimension of the feature vector; M-step (maximization step): using the calculated... Update model parameters: , , , Due to the use of spherical constraints, the variance update formula simplifies to: , Repeat steps E and M, calculating the overall log-likelihood function after each iteration: , Where θ represents all the parameters to be estimated in the Gaussian mixture model, including the mean, variance, and mixture weights; Step S35: Convergence determination and category output.

[0062] When the change in the log-likelihood function between two consecutive iterations is less than a preset threshold (e.g., 1×10), -3 When the algorithm converges, it is determined that the algorithm has converged. At this point, for each target feature point... Output the probability distribution (i.e., the responsibility value vector) of each Gaussian distribution category. The category with the highest probability is selected as the final identification result for the infrared target, while the complete probability distribution is retained for subsequent uncertainty analysis or manual verification. Since each feature vector in the knowledge base is associated with specific statistical and geometric feature values, each identification result can be traced back to the actual thermal radiation parameters and shape parameters of the target, achieving interpretable category decision-making.

[0063] To verify the effectiveness of the method of this invention, experiments were conducted on multiple infrared target detection and recognition datasets. The overall operating environment was deployed on computing devices equipped with high-performance GPUs (such as NVIDIA RTX series) and deep learning frameworks (such as PyTorch). The experiments used the FLIR_ADAS_v2 infrared dataset and a self-built shortwave infrared ship instance segmentation dataset. The results show that the average detection accuracy (mAP) of this invention on FLIR_ADAS_v2 reaches 87.2%; in the ship segmentation task, the mask prediction accuracy is... The accuracy rate reached 76.8%, demonstrating a strong ability to capture blurred infrared edges. In terms of cluster recognition, this invention adopts a spherical covariance-constrained GMM with an adjusted Rand index (ARI) of 0.083 and a normalized mutual information (NMI) of 0.156. Compared with the use of shared covariance (ARI of only 0.013) or arbitrary covariance, the feature data point cluster boundaries of this method are the clearest, and the omission and misclassification rates under high-dimensional data are minimized.

[0064] The method of this invention extracts multi-channel feature maps and instance segmentation masks from infrared images through a target detection network. Based on these, it extracts statistical and geometric attribute features to construct a multi-dimensional feature knowledge base. Then, it uses a Gaussian mixture model constrained by a spherical covariance matrix for clustering and identification to output the target category. This method explicitly constructs a knowledge base with physical meaning, breaks the black box limitation of deep learning, and realizes interpretable and traceable decision-making. At the same time, the spherical constraint eliminates the correlation of feature noise and improves the anti-interference robustness in complex scenes.

[0065] Example 2: This invention provides an infrared target recognition system based on a multidimensional feature knowledge base and a Gaussian mixture model, used to execute the method described in Example 1. The system includes the following modules: Perception module: The perception module is responsible for extracting the regions of interest (ROIs) of infrared targets from the infrared image and acquiring multi-channel feature maps and instance segmentation masks. Specifically, the perception module calls a pre-trained target detection network (e.g., YOLOv11) to perform forward propagation on the input infrared image, outputting the bounding box coordinates, confidence score, category, and instance segmentation mask for each target. Simultaneously, the perception module uses the RoI Align operation to extract multi-scale feature maps from the network's intermediate layers, which are then cropped and averaged to generate a single-channel grayscale target feature map. This module packages the detection results, mask, and feature map and sends them to the knowledge base construction module.

[0066] Knowledge base construction module: The knowledge base construction module receives grayscale target feature maps and instance segmentation masks from the perception module. This module includes a statistical feature extraction unit and a geometric feature extraction unit. The statistical feature extraction unit performs histogram analysis on the grayscale feature map, calculating grayscale mean, variance, energy, entropy, and statistical quantities such as skewness and kurtosis. The geometric feature extraction unit extracts the contour of the binary mask, calculating seven Hu moments, Fourier descriptors, aspect ratio, roundness, and major axis skewness. All feature values ​​are concatenated into a D-dimensional feature vector and associated with information such as target ID and original detection category, and stored in a structured manner in the knowledge base, which can reside in memory or a persistent database.

[0067] Recognition module: The recognition module reads feature data from the knowledge base and first performs Z-score standardization to eliminate dimensional differences. Then, it constructs a Gaussian mixture model, where the covariance matrix of each Gaussian component is forced to be a spherical covariance matrix. The EM algorithm is used to iteratively optimize the model parameters, updating the mean, mixture weights, and spherical variance according to the formula described in Example 1 during the iteration process. Once the log-likelihood function converges, the recognition module outputs the probability distribution of each target belonging to each category and selects the category corresponding to the highest probability as the final recognition result. The recognition result can be returned to the user interface or downstream applications.

[0068] The system of this invention includes a perception module, a knowledge base construction module, and a recognition module, which are respectively responsible for feature map and mask acquisition, statistical and geometric feature extraction and splicing, and mixture Gaussian model clustering recognition with spherical covariance constraints. The system has a clear modular design, supports interpretability and robust recognition, is easy to deploy in high real-time scenarios such as intelligent driving and infrared reconnaissance, and has strong scalability and engineering practicality.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model, characterized in that, include: The process involves acquiring an infrared image to be processed, using a target detection network to extract the region of interest (ROI) of the target in the infrared image, and outputting the corresponding multi-channel feature map and instance segmentation mask. The statistical attribute features of the target are extracted based on the multi-channel feature map, and the geometric attribute features of the target are extracted based on the instance segmentation mask. The statistical attribute features and the geometric attribute features are concatenated into a feature vector for structured storage, thereby constructing a multi-dimensional feature knowledge base. The feature data in the multidimensional feature knowledge base is input into the Gaussian mixture model, and the covariance matrix of each Gaussian component in the Gaussian mixture model is set as a spherical covariance matrix. The expectation-maximization algorithm is used for iterative optimization, and the target category recognition result is output.

2. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The statistical attribute features are used to characterize the thermal radiation distribution characteristics of the target, including: gray mean, gray variance, gray energy, gray entropy, and histogram invariant moments.

3. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The geometric attributes are used to characterize the shape and contour invariance of the target, including: geometric invariant moments, aspect ratio, Fourier descriptor, roundness, and major axis deflection angle.

4. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The spherical covariance matrix is ​​a diagonal matrix in which the variances of each dimension of the features within the same Gaussian component are equal, so that the clusters corresponding to each Gaussian distribution are spherically distributed in the feature space.

5. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 4, characterized in that, The update of the spherical covariance matrix is ​​as follows: based on the weighted responsibility value corresponding to each Gaussian component, the average of the squared Euclidean distances from the feature data points corresponding to each Gaussian component to the mean vector of that Gaussian component is calculated, and then normalized according to the feature dimension.

6. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The output target category identification result includes: outputting the probability distribution of each target belonging to each Gaussian distribution category, and selecting the category corresponding to the highest probability as the final classification result.

7. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The convergence condition of the expectation maximization algorithm is that the change in the log-likelihood function between two consecutive iterations is less than a preset threshold.

8. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, The multi-channel feature map is cropped from different feature levels through region of interest alignment operation, and then the mean value is taken along the channel dimension to generate a single-channel target feature map. Statistical attribute features are extracted based on the single-channel target feature map.

9. The infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model according to claim 1, characterized in that, Before inputting the feature data from the multidimensional feature knowledge base into the Gaussian mixture model, the feature data is Z-score standardized to eliminate the dimensional differences between statistical and geometric attribute features.

10. A system for implementing the infrared target recognition method based on a multidimensional feature knowledge base and a Gaussian mixture model as described in any one of claims 1-9, characterized in that, include: The perception module is used to extract the region of interest of the target from the infrared image and obtain multi-channel feature maps and instance segmentation masks; The knowledge base construction module extracts statistical attribute features based on the multi-channel feature map, extracts geometric attribute features based on the instance segmentation mask of the region of interest, and concatenates the two to construct a multi-dimensional feature knowledge base. The recognition module is used to input feature data from the multidimensional feature knowledge base into a Gaussian mixture model for cluster recognition, wherein the covariance matrix of each Gaussian component of the Gaussian mixture model is constrained to a spherical covariance matrix, and the target's class probability is output.

Images