Tank weakness detection method based on image semantic matching

Abstract

The invention discloses a target detection method based on an image semantic matching algorithm, and the method comprises the steps: carrying out the multi-source data collection through an image sensor, an SAR (Synthetic Aperture Radar) radar and infrared remote sensing equipment which are carried by an unmanned plane, and obtaining a visible light image, a high-resolution radar image and thermal imaging data in a low-light environment; carrying out target identification on different data sources by utilizing an integrated learning method; judging whether the target is a tank or not by adopting a soft voting fusion strategy; after the tank is detected, the crossing unmanned aerial vehicle is dispatched to adjust a flight path, and image data with higher resolution are obtained; a deep learning model is trained based on an image semantic matching technology, semantic analysis is performed on a high-precision image, and weak parts of a target are automatically identified and labeled by matching standard part features. According to the method, the multi-modal source data is fused, the reliability of tank recognition is improved, the weak part of the target is accurately positioned by combining the image semantic matching technology, the method can adapt to a complex battlefield environment, and the intelligent level of battlefield intelligence analysis is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence, computer vision, ensemble learning and multimodal data fusion, and in particular to a tank weakness detection method based on image semantic matching. Background Art

[0002] With the rapid development of drone technology and deep learning in recent years, drones are increasingly being used in military reconnaissance, battlefield monitoring, and target identification. With their flexible maneuverability, low cost, and adaptability to complex terrain, drones have become a vital tool for gathering combat information. In these applications, efficiently and accurately identifying and locating targets from collected information, especially in complex battlefield environments, is crucial for improving combat effectiveness.

[0003] Traditional target recognition methods mostly rely on static images or traditional sensor data, such as optical and infrared images. While these methods can provide some recognition capabilities in specific environments, they are often limited by significant environmental influences (such as weather and lighting conditions), or by camouflaged targets in combat scenarios. In recent years, SAR (Synthetic Aperture Radar) and infrared remote sensing technologies have become important data sources due to their advantages in harsh environments. SAR radar can penetrate clouds and operate at night, while infrared remote sensing technology can identify targets in low-light environments by capturing heat sources. However, the limitations of a single data source make traditional recognition methods difficult to achieve sufficient accuracy and robustness, especially when it is necessary to precisely identify targets and identify their weak points. To improve recognition accuracy and detailed analysis, integrating multiple data sources and combining them with advanced algorithms has become a solution.

[0004] Furthermore, traditional technologies for detecting weak spots in tank images typically rely on expert-designed feature extraction, template matching, region segmentation, and rule-based analysis. These methods often perform well in specific scenarios, but their performance is often limited when faced with changing environments, varying viewpoints, and complex backgrounds, making them difficult to adapt to dynamically changing real-world situations. Summary of the Invention

[0005] Purpose of the invention: The present invention proposes a tank weakness detection method based on image semantic matching to solve the problems of the existing technology in detecting weak points in tank images, such as difficulty in identification and high environmental requirements.

[0006] Technical solution: The tank weakness detection method based on image semantic matching described in the present invention specifically includes the following steps:

[0007] (1) Obtain multimodal data based on the image sensor, SAR radar, and infrared remote sensing equipment carried by the UAV; the image sensor collects visible light images, the SAR radar provides high-resolution radar images, and the infrared remote sensing equipment obtains thermal imaging data;

[0008] (2) Build independent classification models for the collected multi-source data, and use ensemble learning methods to fuse them. Calculate the comprehensive confidence of each category based on the soft voting strategy to identify the type and specific model of the target;

[0009] (3) training the classification model constructed in step (2) using multimodal data, establishing classification performance evaluation indicators, and evaluating the training results;

[0010] (4) When a tank target is detected, the drone’s path is adaptively adjusted to capture high-precision images;

[0011] (5) Based on the acquired high-precision target images, an image semantic matching model is constructed, image features are extracted using deep learning technology, and a cross-image matching strategy is combined to ensure high-precision semantic correspondence of the tank's weak points under different viewing angles;

[0012] (6) Train the image semantic matching model, establish weak point identification evaluation indicators, and evaluate the training results.

[0013] Furthermore, the implementation process of building an independent classification model in step (2) is as follows:

[0014] Independent classification models are constructed for multi-source data. All models adopt a unified deep neural network architecture, consisting of a feature extraction backbone network and a classification head, and are differentiated in hyperparameter settings according to the imaging characteristics of different modalities. The feature extraction backbone network adopts the ResNet deep learning model to extract local and global features from data of different modalities. The classification head consists of two fully connected layers. The first fully connected layer is used to perform nonlinear transformation on the extracted deep features and uses ReLU activation. The second fully connected layer maps the features to the classification category space and calculates the probabilities of tank and non-tank categories through the Softmax function. The model training adopts the cross-entropy loss function and is optimized using the adaptive learning rate optimization algorithm SGD to improve training stability.

[0015] Furthermore, the soft voting strategy is used to calculate the comprehensive confidence of each category in step (2), and the type and specific model of the target are determined as follows:

[0016] Each classification model is initialized using pre-trained weights, and the cross-entropy loss function is used to optimize the classification accuracy. Focal Loss is also used to address the class imbalance problem to enhance the recognition ability of low-confidence targets. The output results of different modal classification models are integrated through a soft voting fusion strategy to calculate the weighted confidence of each category:

[0017]

[0018] Among them, Pfinal(y) represents the confidence of the final target category y, Pm(y) is the confidence of the output of the m-th modal classification model, and wm is the weighting coefficient of the m-th classification model. A dynamic confidence threshold is set to ensure the accuracy of target recognition in complex environments. Based on the classification result, it is determined whether the target is a tank. If a tank target is detected, the subsequent processing flow is entered. If not, data optimization and feature enhancement are continued.

[0019] Furthermore, the implementation process of step (3) is as follows:

[0020] The classification models were trained independently using visible light, SAR radar, and infrared remote sensing data. Each model was optimized using supervised learning. K-fold cross-validation was used during training to evaluate classification performance, and hyperparameters were adjusted based on model performance.

[0021] Data cleaning and data enhancement strategies are used to improve the model's adaptability to different environments and targets. Data enhancement includes random cropping, brightness adjustment, noise interference and geometric transformation; and transfer learning is used to accelerate model convergence. During the training process, the classification performance is evaluated by accuracy, precision, recall rate, F1 score and mean average precision, and the robustness of the model is tested in combination with different weather and lighting conditions. During the optimization process, the cross-entropy loss function is used in combination with the gradient descent method to adjust the model parameters to ensure that the classification model achieves the best performance under multimodal data fusion.

[0022] Furthermore, the implementation process of step (4) is as follows:

[0023] When the target is identified as a tank, the tank's spatial coordinates in the image are calculated, and the drone's flight trajectory is adjusted using a drone path planning algorithm to bring it into the optimal viewing angle. Based on real-time feedback from target detection, the drone's flight parameters are dynamically adjusted to ensure the best resolution image of the tank target is obtained without affecting the battlefield environment.

[0024] After the acquisition is completed, the high-precision images are subjected to denoising, contrast enhancement and edge enhancement processing to provide optimized input data for subsequent tank part labeling.

[0025] Furthermore, the weak parts of the tank described in step (5) include the top of the turret, the tracks, the engine compartment, and the rear of the vehicle.

[0026] Furthermore, the implementation process of step (5) is as follows:

[0027] Feature extraction and normalization: Use a pre-trained ResNet deep convolutional neural network to extract features from the input image, obtaining feature representations Fs and Ft of the standard template image and the image to be matched, respectively. Normalize the feature vectors to keep them within the same scale range to eliminate the influence of different image distributions. After normalization, the feature vectors are used to calculate the pixel-level matching score.

[0028] Feature point matching and similarity calculation: In the matching stage, the degree of matching is measured by calculating the cosine similarity between feature points, which is defined as follows:

[0029]

[0030] Where Cij represents the similarity matching score between the i-th feature point in the standard template image and the j-th feature point in the image to be matched; Fs(i) and Ft(j) represent the feature vectors of the standard template and the image to be matched at the corresponding pixel positions, respectively; ||Fs(i)|| and ||Ft(j)|| represent the Euclidean norm of the feature vector, which is used for normalization calculation;

[0031] Global pixel correspondence modeling: To ensure the consistency of matching structures, a Transformer-based attention mechanism is further used to build global matching relationships. Long-distance information modeling is used to improve the accuracy of component matching under different perspectives and scales. A matching loss function Lmatch is constructed to optimize the matching effect, which is defined as follows:

[0032]

[0033] Among them, the first term ensures that the matching point pairs with high similarity have a small matching error; the second term is a smoothing regularization term, which limits the change in the matching scores of adjacent pixels and ensures the consistency of the matching results in the local area;

[0034] Region segmentation and geometric compensation: After matching is completed, weakly supervised learning methods are used to segment the weak parts of the tank and align the detected parts with the standard template. The "spatial transformation network" is further used to perform geometric transformation compensation to eliminate matching errors caused by perspective changes and improve cross-perspective matching accuracy. Based on the tank weakness feature database, the matched weak parts are highlighted.

[0035] Furthermore, the image semantic matching model is trained in step (6) by using a method combining supervised learning and weakly supervised learning.

[0036] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are: the present invention utilizes multiple sensors carried by drones to construct target classification models suitable for visible light images, SAR images and infrared data based on multi-source data, thereby solving the problem of performance degradation under dynamic changes such as complex environments, different perspectives, occlusion and low light; the present invention innovatively adopts the perspective of image semantic matching to solve the problem of tank weakness detection, and designs a tank part labeling model based on image semantic matching, thereby realizing accurate identification and labeling of tank weak parts. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flowchart of drone operations;

[0038] Figure 2 Flowchart for training image semantic matching model. DETAILED DESCRIPTION

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] The present invention proposes a tank weakness detection method based on image semantic matching, which includes the following steps:

[0041] Step 1, such as Figure 1 As shown, the visible light image sensor, SAR radar and infrared remote sensing equipment carried by the UAV are used to obtain multimodal data. Among them, the visible light image sensor obtains visible light images, the SAR radar obtains high-resolution radar images that can penetrate obstructions, and the infrared remote sensing equipment obtains thermal imaging data in low-light environments to enhance the full-time and space perception capability of tank targets.

[0042] The acquired multimodal data is preprocessed, including denoising, normalization, contrast enhancement, histogram equalization, and pseudo-color conversion, to improve the recognizability of target features and align the spatial distribution of data of different modalities. Based on the labeled tank dataset, data enhancement methods (such as rotation, scaling, flipping, noise perturbation, etc.) are used to improve the generalization ability of the model and ensure that multimodal data can be effectively matched under different lighting, weather conditions, and angles.

[0043] Step 2: Build independent classification models for the collected multi-source data, and use ensemble learning methods to fuse them. Calculate the comprehensive confidence of each category based on the soft voting strategy to identify the type and specific model of the target.

[0044] Since visible light images, SAR images, and infrared data are all stored in two-dimensional matrices, the underlying consistency of their data formats enables the model to extract spatial features through the same convolution operation. ResNet-based classification models for visible light images, SAR radar images, and infrared remote sensing data were constructed for each modality. These models all utilize a unified deep neural network architecture, consisting of a feature extraction backbone network and a classification head. Hyperparameter settings, such as input preprocessing and channel dimension adjustment, are differentiated based on the imaging characteristics of each modality (e.g., speckle noise suppression for SAR and thermal radiation calibration for infrared). The feature extraction backbone network utilizes a ResNet deep learning model to extract local and global features from data from different modalities, thereby enhancing target recognition in complex environments. The classification head consists of two fully connected layers. The first layer performs a nonlinear transformation on the extracted deep features and uses ReLU activation. The second layer maps the features into a classification space and calculates the probabilities of tank and non-tank classes using a softmax function. The model training adopts the cross-entropy loss function (Cross-Entropy Loss) and is optimized using the adaptive learning rate optimization algorithm SGD to improve training stability.

[0045] An ensemble learning method is used to fuse the classification results of the three modal data. The output of the different modal classification models is integrated through a soft voting fusion strategy, and the weighted confidence of each category is calculated. A dynamic confidence threshold is set to ensure accurate target recognition in complex environments.

[0046]

[0047] Among them, Pfinal(y) represents the confidence of the final target category y, Pm(y) is the confidence of the output of the m-th modal classification model, and wm is the weighted coefficient of the m-th classification model (which can be learned through Bayesian optimization or genetic algorithm).

[0048] Based on the classification results, it is determined whether the target is a tank. If a tank target is detected, the subsequent processing flow is entered. If not, data optimization and feature enhancement are continued.

[0049] Step 3: Train the classification model constructed in step 2 using a multimodal dataset, establish classification performance evaluation indicators, and evaluate the training results.

[0050] The classification models constructed in step 2 were independently trained using visible light, SAR radar, and infrared remote sensing data. Each model was optimized using supervised learning. K-fold cross-validation was used during training to evaluate classification performance, and hyperparameters such as batch size and dropout rate were adjusted based on model performance. Each classification model was initialized using pretrained weights, and a cross-entropy loss function was used to optimize classification accuracy. Focal Loss was also used to address class imbalance and enhance the ability to recognize low-confidence targets.

[0051] Data cleaning and augmentation strategies are employed to improve the model's adaptability to diverse environments and targets. Data augmentation includes random cropping, brightness adjustment, noise perturbation, and geometric transformations. Transfer learning is also used to accelerate model convergence. During training, classification performance is evaluated using metrics such as accuracy, precision, recall, F1 score, and mean average precision (mAP). Model robustness is also tested under varying weather and lighting conditions. During optimization, a cross-entropy loss function is used in conjunction with gradient descent to adjust model parameters, ensuring optimal performance of the classification model using multimodal data fusion.

[0052] Step 4: The drone performs adaptive path adjustment and high-precision image acquisition to detect and identify weak points of the tank, thereby improving the accuracy of subsequent processing.

[0053] When target recognition identifies a tank, the system calculates the tank's spatial coordinates within the image and, using a drone path planning algorithm (reinforcement learning), adjusts the drone's flight trajectory to the optimal viewing angle. Based on real-time feedback from target detection, the system dynamically adjusts drone flight parameters (such as speed, altitude, and gimbal angle) to ensure optimal resolution of the tank without compromising the battlefield environment. After acquisition, the high-precision image undergoes denoising, contrast enhancement, and edge enhancement, providing optimized input data for subsequent tank part annotation.

[0054] Step 5: Based on the high-precision tank image obtained in step 4, an image semantic matching model is constructed to automatically identify the weak points of the tank and accurately mark the potential weak points. Deep learning technology is used to extract image features and combined with cross-image matching strategies to ensure that high-precision semantic correspondence of tank parts can be achieved under different viewing angles. The model training process is as follows: Figure 2 shown.

[0055] Based on an existing database of standard tank structures, we constructed a high-quality dataset of standard tank part templates. This dataset covers images of tanks of different models and angles, and includes manual annotations of vulnerable areas such as turrets, tracks, and wheels. The data is formatted as tank image pairs, consisting of an original tank image A and an annotated tank image B.

[0056] Feature extraction and normalization. A pretrained ResNet deep convolutional neural network is used to extract features from the input image, obtaining the feature representations Fs and Ft of the standard template tank image B and the tank image to be matched, respectively. To ensure robust matching, the feature vectors are normalized to the same scale range to eliminate the influence of different image distributions. After normalization, the feature vectors are used to calculate the pixel-level matching score.

[0057] Feature point matching and similarity calculation. In the matching stage, the degree of matching is measured by calculating the cosine similarity between feature points, which is defined as follows:

[0058]

[0059] Among them, Cij represents the similarity matching score between the i-th feature point in the standard template image and the j-th feature point in the image to be matched; Fs(i) and Ft(j) represent the feature vectors of the standard template and the image to be matched at the corresponding pixel positions, respectively; ||Fs(i)|| and ||Ft(j)|| represent the Euclidean norm (i.e., L2 norm) of the feature vector, which is used for normalization calculation.

[0060] Global pixel correspondence modeling. To ensure consistency in matching structures, a Transformer-based attention mechanism is further employed to construct global matching relationships. This method improves component matching accuracy across different viewpoints and scales by modeling long-range information. Based on this, a matching loss function, Lmatch, is constructed to optimize matching performance, defined as follows:

[0061]

[0062] The first term ensures that highly similar matching point pairs have a small matching error. The second term is a smoothing regularization term that limits the variation in matching scores between adjacent pixels, ensuring consistent matching results in local areas. Based on this loss function, the ultimate goal is to train a pixel-wise matching score matrix between the standard template tank image B and the tank image A to be matched.

[0063] Furthermore, the class activation map is introduced into the matching matrix optimization process. The pixel-level classification results generated by the semantic segmentation network are used to subdivide the image pixels into four semantic levels: hotspots (focused areas), main objects, contextual regions, and irrelevant background regions. Based on this hierarchical mechanism, a hierarchical weight constraint is constructed, in which hotspots are assigned the highest weight coefficients. This structured weight allocation strategy strengthens the feature response of key areas while suppressing the interference of background noise, allowing the matching process to better focus on visual elements with significant semantic information.

[0064] Region segmentation and geometric compensation. After matching is complete, weakly supervised learning methods are used to segment the tank's weak areas and align the detected areas with a standard template. A "spatial transformation network" is further used to perform geometric transformation compensation to eliminate matching errors caused by perspective changes, thereby improving cross-viewpoint matching accuracy.

[0065] A high-precision tank image is input, and based on the semantic matching model, the pixel-level semantic correspondence between the image and the standard template is calculated, and a matching score matrix is output. Optical flow estimation (RAFT), correlation matching (DINO), or regularized Hough transform are used to optimize the spatial alignment of weak parts such as the tank turret, tracks, and wheels, and part annotation information is automatically generated based on the matching results. Combined with the weakness recognition database, the weak area is annotated with high brightness in the output image according to the annotated parts. Finally, the annotated tank image is output and tank weakness distribution data is generated to provide auxiliary decision support for battlefield intelligence analysis and weapon strikes.

[0066] Step 6: Train the image semantic matching model constructed in step 5, build a large number of target feature annotation data sets for training, establish weak position recognition evaluation indicators, and evaluate the training results.

[0067] During model training, a combination of supervised and weakly supervised learning is used to ensure that the model can effectively learn the matching rules for target parts even when some labeled data is missing. A high-quality tank dataset was constructed, including RGB images from different angles and environments, with detailed annotations of vulnerable areas (such as the turret, tracks, engine compartment, and vehicle rear) for training the semantic matching model. Weak point identification evaluation metrics, including matching accuracy, recall rate, and weakness prediction accuracy, were constructed based on factors such as tank thickness and structural fragility. The training results were evaluated to ensure the accuracy and robustness of the model.

[0068] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0069] The present invention has been described in detail above with reference to specific embodiments. However, these descriptions should not be construed as limiting the present invention. Those skilled in the art will appreciate that various equivalent substitutions, modifications, or improvements may be made to the technical solutions and implementations of the present invention without departing from the spirit and scope of the present invention, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A tank weakness detection method based on image semantic matching, characterized in that: The following steps are involved: (1) Obtain multimodal data based on the image sensor, SAR radar, and infrared remote sensing equipment carried by the UAV; the image sensor collects visible light images, the SAR radar provides high-resolution radar images, and the infrared remote sensing equipment obtains thermal imaging data; (2) Build independent classification models for the collected multi-source data, and use ensemble learning methods to fuse them. Calculate the comprehensive confidence of each category based on the soft voting strategy to identify the type and specific model of the target; (3) training the classification model constructed in step (2) using multimodal data, establishing classification performance evaluation indicators, and evaluating the training results; (4) When a tank target is detected, the drone’s path is adaptively adjusted to capture high-precision images; (5) Based on the acquired high-precision target images, an image semantic matching model is constructed, image features are extracted using deep learning technology, and a cross-image matching strategy is combined to ensure high-precision semantic correspondence of the tank's weak points under different viewing angles; (6) Train the image semantic matching model, establish weak point identification evaluation indicators, and evaluate the training results.

2. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The implementation process of building an independent classification model in step (2) is as follows: Independent classification models are constructed for multi-source data. Each model uses a unified deep neural network architecture, consisting of a feature extraction backbone network and a classification head. Hyperparameter settings are differentiated to suit the imaging characteristics of different modalities. The feature extraction backbone network uses the ResNet deep learning model to extract local and global features from data of different modalities. The classification head consists of two fully connected layers. The first fully connected layer is used to perform nonlinear transformation on the extracted deep features and uses ReLU activation. The second fully connected layer maps the features to the classification category space and calculates the probabilities of tank and non-tank categories through the Softmax function. The model training adopts the cross-entropy loss function and is optimized using the adaptive learning rate optimization algorithm SGD to improve training stability.

3. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The implementation process of calculating the comprehensive confidence of each category based on the soft voting strategy in step (2) and distinguishing the type and specific model of the target is as follows: Each classification model is initialized using pre-trained weights, and the cross-entropy loss function is used to optimize the classification accuracy. Focal Loss is also used to address the class imbalance problem to enhance the recognition ability of low-confidence targets. The output results of different modal classification models are integrated through a soft voting fusion strategy to calculate the weighted confidence of each category: Among them, Pfinal(y) represents the confidence of the final target category y, Pm(y) is the confidence of the output of the m-th modal classification model, and wm is the weighting coefficient of the m-th classification model. A dynamic confidence threshold is set to ensure the accuracy of target recognition in complex environments. Based on the classification result, it is determined whether the target is a tank. If a tank target is detected, the subsequent processing flow is entered. If not, data optimization and feature enhancement are continued.

4. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The implementation process of step (3) is as follows: The classification models were trained independently using visible light, SAR radar, and infrared remote sensing data. Each model was optimized using supervised learning. K-fold cross-validation was used during training to evaluate classification performance, and hyperparameters were adjusted based on model performance. Data cleaning and data augmentation strategies are used to improve the model's adaptability to different environments and targets. Data augmentation includes random cropping, brightness adjustment, noise interference, and geometric transformation. Transfer learning is also used to accelerate model convergence. During the training process, the classification performance is evaluated by accuracy, precision, recall, F1 score and mean average precision, and the model robustness is tested in combination with different weather and lighting conditions. During the optimization process, the cross-entropy loss function is used in combination with the gradient descent method to adjust the model parameters to ensure that the classification model achieves the best performance under multimodal data fusion.

5. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The implementation process of step (4) is as follows: When the target is identified as a tank, the tank's spatial coordinates in the image are calculated, and the drone's flight trajectory is adjusted using a drone path planning algorithm to bring it into the optimal viewing angle. Based on real-time feedback from target detection, the drone's flight parameters are dynamically adjusted to ensure the best resolution image of the tank target is obtained without affecting the battlefield environment. After the acquisition is completed, the high-precision images are subjected to denoising, contrast enhancement and edge enhancement processing to provide optimized input data for subsequent tank part labeling.

6. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The weak parts of the tank described in step (5) include the top of the turret, the tracks, the engine compartment, and the rear of the vehicle.

7. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The implementation process of step (5) is as follows: Feature extraction and normalization: Use a pre-trained ResNet deep convolutional neural network to extract features from the input image, obtaining feature representations Fs and Ft of the standard template image and the image to be matched, respectively. Normalize the feature vectors to keep them within the same scale range to eliminate the influence of different image distributions. After normalization, the feature vectors are used to calculate the pixel-level matching score. Feature point matching and similarity calculation: In the matching stage, the degree of matching is measured by calculating the cosine similarity between feature points, which is defined as follows: Where Cij represents the similarity matching score between the i-th feature point in the standard template image and the j-th feature point in the image to be matched; Fs(i) and Ft(j) represent the feature vectors of the standard template and the image to be matched at the corresponding pixel positions, respectively; ||Fs(i)|| and ||Ft(j)|| represent the Euclidean norm of the feature vector, which is used for normalization calculation; Global pixel correspondence modeling: To ensure the consistency of matching structures, a Transformer-based attention mechanism is further used to build global matching relationships. Long-distance information modeling is used to improve the accuracy of component matching under different perspectives and scales. A matching loss function Lmatch is constructed to optimize the matching effect, which is defined as follows: Among them, the first term ensures that the matching point pairs with high similarity have a small matching error; the second term is a smoothing regularization term, which limits the change in the matching scores of adjacent pixels and ensures the consistency of the matching results in the local area; Region Segmentation and Geometric Compensation: After matching is complete, weakly supervised learning methods are used to segment the tank's weak points and align the detected areas with a standard template. A "spatial transformation network" is further used to perform geometric transformation compensation to eliminate matching errors caused by perspective changes and improve cross-viewpoint matching accuracy. Based on the tank weakness feature database, the matched weak points are highlighted.

8. The tank weakness detection method based on image semantic matching according to claim 1 is characterized in that: The image semantic matching model is trained in step (6) by combining supervised learning and weakly supervised learning.

Images