A transfer learning-driven method for UXO crater location and identification

Abstract

The present invention discloses a method for locating and identifying unexploded bomb craters driven by transfer learning. First, a data set is constructed, and remote sensing images and drone aerial images are respectively matched to source domain and target domain data sets; the source domain data set is used to train the basic model, and the balanced data set is used to train the migration model. Finally, the model is inferred. If the confidence score of the target exceeds the threshold, the target is considered to be credible and there may be unexploded bombs. The present invention uses drone images and deep learning technology to quickly and accurately locate and identify unexploded bomb craters, improves the efficiency and accuracy of the unexploded bomb search process, and provides strong support for battlefield cleaning and reconstruction.

Description

Technical Field

[0001] The invention belongs to the technical field of pattern recognition, and in particular relates to a method for locating and identifying an unexploded bomb crater. Background Art

[0002] Unexploded bombs refer to unexploded or destroyed weapons and ammunition in war-torn areas, conflict zones or military exercise areas. They may cause accidental explosions, causing serious harm and losses to people and social property, and posing a potential safety hazard. At present, the positioning and identification of unexploded bombs mainly rely on manual removal or the use of metal detectors, geological radars and other equipment for detection, which is inefficient and highly dangerous. Therefore, it is of great significance to automatically locate and accurately identify unexploded bombs.

[0003] Drone images have the advantages of high resolution, multi-angle observation, instant acquisition, and low cost, making them suitable for searching for unexploded bombs. Craters are the most direct evidence of the existence of unexploded bombs and are also the most common war-related objects in aerial images. Therefore, the search for unexploded bombs can also be defined as the detection of unexploded bomb craters. Once the craters are detected, the location where the unexploded bombs may exist can be inferred. At present, there are two main problems in the positioning and identification of unexploded bomb craters. First, the number of image data of unexploded bomb craters is insufficient. Due to its scarcity, the number of samples in certain categories may be limited, resulting in an unbalanced category distribution. Second, unexploded bomb craters may be similar to the background in appearance, texture, or color, which can easily lead to false detection. The key to solving these problems is to improve the generalization performance of the model so as to accurately locate and identify unexploded bomb craters. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a method for locating and identifying unexploded bomb craters driven by transfer learning. First, a data set is constructed, and remote sensing images and drone aerial images are respectively matched to source domain and target domain data sets; the source domain data set is used to train the basic model, and the balanced data set is used to train the migration model. Finally, the model is inferred. If the confidence score of the target exceeds the threshold, the target is considered to be credible and there may be unexploded bombs. The present invention uses drone images and deep learning technology to quickly and accurately locate and identify unexploded bomb craters, improves the efficiency and accuracy of the unexploded bomb search process, and provides strong support for battlefield cleaning and reconstruction.

[0005] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0006] Step 1: Dataset construction;

[0007] Obtain high-resolution remote sensing images / drone aerial images of war-torn areas, conflict zones, or military exercise areas, screen and preprocess the images, manually annotate craters, and divide them into training and test sets;

[0008] Remote sensing images and drone aerial images correspond to the source domain and target domain datasets respectively;

[0009] Step 2: Train the basic model with the source domain dataset;

[0010] The basic model includes a feature extraction network, a feature pyramid, a ROI pooling layer, a ROI feature extraction network, a region generation network, a classifier and a regressor;

[0011] The source image is input into the feature extraction network, and then passes through the feature pyramid, ROI pooling layer, and ROI feature extraction network in sequence; the feature pyramid is also connected to the ROI pooling layer through the region generation network;

[0012] The feature extraction network extracts features from the image;

[0013] The feature pyramid fuses multi-scale context information;

[0014] The region generation network generates candidate regions by sliding a window on the feature map and outputs confidence scores and position offset information;

[0015] The ROI pooling layer maps candidate regions of different sizes into feature representations of fixed dimensions;

[0016] The ROI feature extraction network uses a convolutional layer to further extract features from the results of the ROI pooling layer; the extracted features are then sent to the classifier and the regressor respectively;

[0017] The classifier and regressor classify and regress features through a fully connected layer;

[0018] The base model is trained with a joint loss function, including the loss of the region generation network, the cross entropy loss of the classifier, and the smoothed L1 loss of the regressor;

[0019] Iteratively optimize the model parameters through the gradient descent algorithm until the requirements are met or the model accuracy is no longer improved;

[0020] Step 3: Balance the dataset to train the migration model;

[0021] Randomly sample an equal number of source domain images and target domain images to construct a balanced dataset;

[0022] The migration model adds a contrast representation embedding branch to the basic model; in the migration model, the features extracted by the ROI feature extraction network are respectively sent to the classifier, the regressor and the contrast representation embedding branch; the classifier of the migration model uses a cosine similarity classifier;

[0023] The cosine similarity classifier performs classification judgment by measuring the similarity between the candidate region features and the class weight vector; the contrast representation embedding branch uses a multi-layer perceptron structure to encode the features of the candidate region into a contrast representation embedding vector, and designs a loss function to make samples of the same category more compact in the feature space and samples of different categories gradually move away;

[0024] The migration model fixes the parameters of the feature extraction network during the fine-tuning process, and optimizes the parameters of other modules through the gradient descent algorithm until the model accuracy requirements are met or no longer improved.

[0025] Step 4: Model inference;

[0026] Given a set of images to be tested, a set of prediction results are obtained through the trained model. Each result contains the bounding box coordinates, predicted category and confidence score of the target. A confidence threshold δ is set. If the confidence score of the target exceeds the threshold, the target is considered to be credible and there may be unexploded bombs.

[0027] Preferably, the step 1 is specifically:

[0028] Source domain dataset construction: obtain high-resolution remote sensing images of war-torn areas, conflict areas, or military exercise areas, and screen images containing craters; use ENVI software to preprocess the screened images, including correcting geometric distortion and eliminating atmospheric interference; cut the preprocessed images into fixed-size image blocks, annotate them, add labels, draw bounding boxes, and clearly mark the craters in the images; divide the training set and test set into an 8:2 ratio to obtain the final source domain dataset;

[0029] Construction of the target domain dataset: Use drone equipment to take aerial photos of war-torn areas, conflict areas, or military exercise areas to obtain high-resolution aerial images; filter out images containing craters for annotation, add labels, and draw bounding boxes; divide the training set and test set into an 8:2 ratio to obtain the final target domain dataset.

[0030] Preferably, the step 2 is specifically:

[0031] The feature extraction network uses a deep residual convolution structure to extract features from the input image;

[0032] The feature pyramid constructs and fuses multi-scale feature maps to extract semantic and detail information at different levels;

[0033] The region generation network moves on the feature map in a sliding window manner, generates candidate regions of different sizes and proportions for each window position, and performs target / background classification and bounding box regression through a fully connected layer; outputs the confidence score and position offset information of each candidate region, selects candidate regions with high confidence and inputs them into subsequent classification and regression networks for target detection and positioning;

[0034] The ROI pooling layer maps the candidate region feature maps of different sizes into feature representations of fixed dimensions;

[0035] The ROI feature extraction network uses a convolutional layer to further extract features from the results of the ROI pooling layer;

[0036] The classifier and regressor are composed of two fully connected layers; for each candidate region, the classifier predicts the probability that it belongs to different target categories, and the regressor predicts the coordinate offset of its bounding box;

[0037] In the basic model training stage, the joint loss function is expressed as

[0038]

[0039] in Acting on the output of the region generation network to distinguish foreground from background and refine the bounding box; is the cross entropy loss for the classifier; is the smoothed L1 loss for the regressor.

[0040] Preferably, the step 3 is specifically:

[0041] Randomly sample equal numbers of source domain images and target domain images to construct a balanced dataset; the number of samples is set to the total number of samples in the target domain dataset, that is, all target domain data is added to the balanced dataset;

[0042] The transfer model uses a cosine similarity-based classifier with an additional contrast representation embedding branch; the weight matrix of the classifier It is expressed as [ω1,ω2,…,ω c ],in is the class weight vector for each category; the output of the classifier is the similarity score P between the input feature x and each class weight vector:

[0043]

[0044] Among them, p i,j is the similarity score between the i-th candidate region of x and the weight vector of category j, and α represents the scaling factor;

[0045] The contrastive representation embedding branch adopts a multi-layer perceptron structure to encode ROI features into contrastive representation embedding vectors. By measuring the similarity scores between different candidate regions, a loss function is designed to make samples of the same category more compact in the feature space, while samples of different categories gradually move away from each other in the feature space:

[0046]

[0047]

[0048] where z i is the contrast representation embedding vector of the i-th ROI feature encoding, u i Represents the intersection over union score IoU,y with the matching bounding box label i Represents the category label; Yes and y i The number of candidate regions of the same category, τ is the scaling factor; represents the normalized eigenvector, Measures the cosine similarity between the i-th and j-th candidate regions; f(u i ) to filter the candidate areas:

[0049]

[0050] where g(·) assigns different weight coefficients to objects with different IoU scores; φ is the IoU threshold;

[0051] The joint loss function in the fine-tuning stage is expressed as:

[0052]

[0053] where λ is The weight factor of

[0054] The migration model is initialized using the basic model parameters obtained in step 2. During fine-tuning, the parameters of the feature extraction network are fixed, and the parameters of the remaining modules are optimized through iterative training of the gradient descent algorithm until the model accuracy no longer improves or meets the requirements.

[0055] Preferably, φ is set to 0.7, and λ is set to 0.5.

[0056] Preferably, the step 4 is specifically:

[0057] Given a set of test images, input the images into the model trained in step 3 to obtain a set of prediction results {(x1 i ,y1 i ,x2 i ,x2 i ,c i ,pi )|i=1,…,M}, where (x1 i ,y1 i ,x2 i ,x2 i ) is the bounding box coordinate of the i-th detected object, c i and p i Represent the corresponding prediction category and confidence score respectively;

[0058] The confidence threshold is set to δ. If the confidence score of a target is higher than the threshold, the target is considered to be credible, that is, there may be unexploded bombs.

[0059] The beneficial effects of the present invention are as follows:

[0060] The transfer learning method proposed in this paper makes comprehensive use of the advantages of remote sensing images and drone images, effectively solves the problem of insufficient target domain data and distribution differences, and improves the accuracy of unexploded bomb crater positioning and identification. The cosine similarity classifier can provide accurate classification discrimination by measuring the similarity between the candidate region features and the category weight vector; the contrast representation embedding branch designs a specific distance measurement method, so that samples of the same category are more compactly clustered in the feature space, while samples of different categories gradually move away, thereby effectively alleviating the interference of similar targets in the background. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the overall structure of the model in the method of the present invention. DETAILED DESCRIPTION

[0062] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0063] The present invention discloses a method for locating and identifying unexploded bomb craters driven by transfer learning. The method constructs a high-resolution remote sensing image and drone image dataset, and aims to solve the problems of insufficient data volume and uneven distribution through transfer learning technology, thereby improving the generalization performance of locating and identifying unexploded bomb craters. At the same time, by comparing the distance between the candidate area features and the known target features, it can effectively distinguish targets with similar appearance features in the background, reducing the risk of misjudgment.

[0064] The technical solution includes the following steps:

[0065] 1. Dataset construction. Obtain high-resolution remote sensing images / drone aerial images of war-torn areas, conflict areas, or military exercise areas, screen and preprocess the images, manually annotate craters, and divide them into training and test sets. Remote sensing images and drone aerial images correspond to source and target domain datasets, respectively.

[0066] 2. Train the basic model with the source domain dataset. The feature extraction network extracts features from the image, the feature pyramid network fuses multi-scale contextual information, and the region generation network generates candidate regions by sliding windows on the feature map, and outputs confidence scores and position offset information. ROI pooling maps candidate region feature maps of different sizes into feature representations of fixed dimensions. The ROI feature extraction network uses convolutional layers to further extract features from the results of ROI pooling, and the classifier and regressor classify and regress these features through fully connected layers. The basic model is trained using a joint loss function, including the loss of the region generation network, the cross entropy loss of the classifier, and the smoothed L1 loss of the regressor. The model parameters are iteratively optimized using the gradient descent algorithm until the requirements are met or the model accuracy no longer improves.

[0067] 3. Train the migration model on a balanced dataset. Randomly sample an equal number of source domain images and target domain images to construct a balanced dataset. The migration model uses a cosine similarity-based classifier and adds a contrastive representation embedding branch. The cosine similarity classifier makes classification decisions by measuring the similarity between the candidate region features and the class weight vector, which enhances the model's generalization ability for new categories. The contrastive representation embedding branch uses a multi-layer perceptron structure to encode the features of the candidate region into a contrastive representation embedding vector, and designs a loss function that makes samples of the same category more compact in the feature space and samples of different categories gradually move away. During the fine-tuning process, the migration model fixes the parameters of the feature extraction network, and optimizes other module parameters (such as classifiers and regressors) through the gradient descent algorithm until the model accuracy requirements are met or no longer improve.

[0068] 4. Model reasoning. Given a set of images to be tested, a set of prediction results are obtained through the trained model. Each result contains the bounding box coordinates, predicted category and confidence score of the target. A confidence threshold δ is set. If the confidence score of the target exceeds the threshold, the target is considered to be credible and there may be unexploded bombs.

[0069] Example:

[0070] Reference Figure 1 The specific implementation steps of the present invention are as follows:

[0071] Step 1: Dataset construction.

[0072] Source domain dataset construction: Obtain high-resolution remote sensing images of war remnants, conflict areas, or military exercise areas, and screen images containing craters. Use ENVI software to preprocess the screened images, including correcting geometric distortion and eliminating atmospheric interference to improve image quality and usability. Cut the preprocessed large images into fixed-size image blocks, perform manual annotation, add labels, draw bounding boxes, and clearly mark the craters in the images. Divide the training set and test set in an 8:2 ratio to obtain the final source domain dataset.

[0073] Construction of target domain dataset: Use drone equipment to take aerial photos of war-torn areas, conflict areas, or military exercise areas to obtain high-resolution aerial images. Screen images containing craters for manual annotation, add labels, or draw bounding boxes. Divide the training set and test set in an 8:2 ratio to obtain the final target domain dataset.

[0074] Step 2: Train the basic model using the source domain dataset.

[0075] The network structure of the basic model is as follows Figure 1 (a). The feature extraction network uses a deep residual convolution structure to extract features from the input image. The feature pyramid network constructs and fuses multi-scale feature maps to extract semantic and detail information at different levels. The region generation network moves on the feature map in a sliding window manner, generates candidate regions of different sizes and proportions for each window position, and performs target / background classification and bounding box regression through a fully connected layer. The region generation network outputs the confidence score and position offset information of each candidate region, and selects candidate regions with high confidence to be input into subsequent classification and regression networks for target detection and positioning. ROI pooling maps the feature maps of candidate regions of different sizes into feature representations of fixed dimensions, and the ROI feature extraction network uses convolutional layers to further extract features from the results of ROI pooling. The classifier and regressor consist of two fully connected layers. For each candidate region, the classifier predicts the probability of it belonging to different target categories, and the regressor predicts the coordinate offset of its bounding box.

[0076] In the basic model training stage, the joint loss function is expressed as

[0077]

[0078] in Acting on the output of the region generation network to distinguish foreground from background and refine the bounding box; is the cross entropy loss for the classifier; is the smoothed L1 loss for the regressor.

[0079] Iterative training is performed through the gradient descent algorithm to optimize the model parameters until the model accuracy no longer improves or meets the requirements.

[0080] Step 3: Balance the dataset and fine-tune the migration model.

[0081] In order to alleviate the imbalance problem in the number and distribution of samples between the source domain and the target domain, an equal number of source domain images and target domain images are randomly sampled to construct a balanced dataset. The number of samples is set to the total number of samples in the target domain dataset, that is, all the target domain data is added to the balanced dataset.

[0082] Different from the base model, the transfer model uses a cosine similarity-based classifier and adds an additional contrast representation embedding branch, such as Figure 1 (b) shows the weight matrix of the classifier. It can be expressed as [ω1,ω2,…,ω c ],in is the class weight vector for each category. The output of the classifier is the similarity score P between the input feature x and each class weight vector:

[0083]

[0084] Among them, p i,j is the similarity score between the i-th candidate region of x and the weight vector of category j, and α represents the scaling factor. Compared with the fully connected layer classifier, the cosine similarity classifier helps to reduce the intra-class variance, has better discrimination ability for samples of new categories, and does not significantly reduce the detection accuracy of the base class.

[0085] The contrastive representation embedding branch uses a multi-layer perceptron structure to encode ROI features into contrastive representation embedding vectors. By measuring the similarity scores between different candidate regions, the loss function is designed to make samples of the same category more compact in the feature space, while samples of different categories gradually move away from each other in the feature space:

[0086]

[0087]

[0088] where z i is the contrast representation embedding vector of the i-th ROI feature encoding, u i Represents the intersection over union score IoU,y with the matching bounding box label i Represents the category label. Yes and y i The number of candidate regions with the same category, τ is the scaling factor. represents the normalized eigenvector, so It measures the cosine similarity between the i-th and j-th candidate regions. Considering that the candidate regions with lower IoU may contain irrelevant semantics, f(u i ) Screen the candidate regions and exclude those with large deviations, thereby improving the model's ability to capture relevant semantics:

[0089]

[0090] where g(·) assigns different weight coefficients to objects with different IoU scores, and φ is the IoU threshold, which is set to 0.7 by default.

[0091] The joint loss function in the fine-tuning stage is expressed as:

[0092]

[0093] where λ is The default setting is 0.5.

[0094] The migration model is initialized using the basic model parameters obtained in step 2. During the fine-tuning process, the parameters of the feature extraction network are fixed, and the parameters of the remaining modules (such as classifiers and regressors) are optimized through iterative training of the gradient descent algorithm until the model accuracy no longer improves or meets the requirements.

[0095] Step 4: Model reasoning.

[0096] Given a set of images to be tested (such as a test set), the images are input into the model trained in step 3 to obtain a set of prediction results {(x1 i ,y1 i ,x2 i ,x2 i ,c i ,p i )|i=1,…,M}, where (x1 i ,y1 i ,x2 i ,x2 i ) is the bounding box coordinate of the i-th detected object, c i and p i Represent the corresponding prediction category and confidence score respectively.

[0097] The confidence threshold is set to δ. If the confidence score of a target is higher than the threshold, the target is considered to be credible, that is, there may be unexploded bombs.

[0098] The effect of the present invention can be further illustrated by the following experiments.

[0099] 1. Experimental conditions

[0100] The present invention is that the central processing unit is The experiment was conducted using Python language with i9-10920X 3.5GHz CPU, 3090 24G GPU, 128G memory and WINDOWS10 operating system.

[0101] The source domain dataset used in the experiment comes from the Gaojing-1 satellite, containing 1280 images and 1720 crater target instances. The target dataset is taken by a DJI drone and contains 75 images and 90 crater target instances. During the training process, the initial learning rate is set to 0.01, the momentum is 0.9, the weight decay is 0.0005, and the batch size is 16. The maximum number of training rounds for the base model and the migration model are 300 and 100, respectively. The feature extraction network selects Resnet-101, and the encoding feature dimension of the contrast representation embedding branch is set to 128.

[0102] 2. Experimental content

[0103] The present invention uses mAP (mean Average Precision) to measure the positioning and recognition accuracy of crater targets, and conducts experiments on a target domain test set constructed based on drone images. The experimental results are shown in Table 1.

[0104] Table 1 Experimental results

[0105]

[0106] The AP of the present invention on the target domain test set 50 and AP 75 are 87.23 and 73.84 respectively, which verifies the effectiveness of the present invention. In general, the present invention realizes the effective migration from remote sensing images to drone images, alleviates the problems of insufficient target domain data and unbalanced category distribution, and enhances the generalization performance of unexploded bomb crater location and identification. Designing a specific distance measurement method and loss function makes the samples within the class more compact and the samples between classes more distant during the training process. On this basis, the interference of similar targets in the background is effectively distinguished by measuring the similarity for classification. The present invention uses drone images and deep learning technology to quickly and accurately locate and identify unexploded bomb craters, improves the efficiency and accuracy of the unexploded bomb search process, and provides strong support for battlefield cleaning and reconstruction.

Claims

1. A transfer learning driven unexploded bomb crater location and identification method, characterized in that: The steps include: Step 1: Dataset construction; Obtain high-resolution remote sensing images / drone aerial images of war-torn areas, conflict zones, or military exercise areas, screen and preprocess the images, manually annotate craters, and divide them into training and test sets; Remote sensing images and drone aerial images correspond to the source domain and target domain datasets respectively; Step 2: Train the basic model with the source domain dataset; The basic model includes a feature extraction network, a feature pyramid, a ROI pooling layer, a ROI feature extraction network, a region generation network, a classifier and a regressor; The source image is input into the feature extraction network, and then passes through the feature pyramid, ROI pooling layer, and ROI feature extraction network in sequence; The feature pyramid is also connected to the ROI pooling layer through the region generation network; The feature extraction network extracts features from the image; The feature pyramid fuses multi-scale context information; The region generation network generates candidate regions by sliding a window on the feature map and outputs confidence scores and position offset information; The ROI pooling layer maps candidate regions of different sizes into feature representations of fixed dimensions; The ROI feature extraction network uses a convolutional layer to further extract features from the results of the ROI pooling layer; the extracted features are then sent to the classifier and the regressor respectively; The classifier and regressor classify and regress features through a fully connected layer; The base model is trained with a joint loss function, including the loss of the region generation network, the cross entropy loss of the classifier, and the smoothed L1 loss of the regressor; Iteratively optimize the model parameters through the gradient descent algorithm until the requirements are met or the model accuracy is no longer improved; Step 3: Balance the dataset to train the migration model; Randomly sample an equal number of source domain images and target domain images to construct a balanced dataset; The migration model adds a contrast representation embedding branch to the basic model; in the migration model, the features extracted by the ROI feature extraction network are respectively sent to the classifier, the regressor and the contrast representation embedding branch; the classifier of the migration model uses a cosine similarity classifier; The cosine similarity classifier performs classification judgment by measuring the similarity between the candidate region features and the class weight vector; the contrast representation embedding branch uses a multi-layer perceptron structure to encode the features of the candidate region into a contrast representation embedding vector, and designs a loss function to make samples of the same category more compact in the feature space and samples of different categories gradually move away; The migration model fixes the parameters of the feature extraction network during the fine-tuning process, and optimizes the parameters of other modules through the gradient descent algorithm until the model accuracy requirements are met or no longer improve; Step 4: Model inference; Given a set of images to be tested, a set of prediction results are obtained through the trained model. Each result contains the bounding box coordinates, predicted category and confidence score of the target. A confidence threshold δ is set. If the confidence score of the target exceeds the threshold, the target is considered to be credible and there may be unexploded bombs.

2. The method for locating and identifying unexploded bomb craters driven by transfer learning according to claim 1, characterized in that: The step 1 is specifically as follows: Source domain dataset construction: obtain high-resolution remote sensing images of war-torn areas, conflict areas, or military exercise areas, and screen images containing craters; use ENVI software to preprocess the screened images, including correcting geometric distortion and eliminating atmospheric interference; cut the preprocessed images into fixed-size image blocks, annotate them, add labels, draw bounding boxes, and clearly mark the craters in the images; divide the training set and test set into an 8:2 ratio to obtain the final source domain dataset; Construction of target domain dataset: Use drone equipment to take aerial photos of war-torn areas, conflict areas, or military exercise areas to obtain high-resolution aerial images; filter images containing craters for annotation, add labels, and draw bounding boxes; The training set and test set are divided into 8:2 ratio to obtain the final target domain dataset.

3. The method for locating and identifying unexploded bomb craters driven by transfer learning according to claim 2, characterized in that: The step 2 is specifically as follows: The feature extraction network adopts a deep residual convolution structure to extract features from the input image; The feature pyramid constructs and fuses multi-scale feature maps to extract semantic and detail information at different levels; The region generation network moves on the feature map in a sliding window manner, generates candidate regions of different sizes and proportions for each window position, and performs target / background classification and bounding box regression through a fully connected layer; outputs the confidence score and position offset information of each candidate region, selects candidate regions with high confidence and inputs them into subsequent classification and regression networks for target detection and positioning; The ROI pooling layer maps the candidate region feature maps of different sizes into feature representations of fixed dimensions; The ROI feature extraction network uses a convolutional layer to further extract features from the results of the ROI pooling layer; The classifier and regressor are composed of two fully connected layers; for each candidate region, the classifier predicts the probability that it belongs to different target categories, and the regressor predicts the coordinate offset of its bounding box; In the basic model training stage, the joint loss function is expressed as in Acting on the output of the region generation network to distinguish foreground from background and refine the bounding box; is the cross entropy loss for the classifier; is the smoothed L1 loss for the regressor.

4. The method for locating and identifying unexploded bomb craters driven by transfer learning according to claim 3, characterized in that: The step 3 is specifically as follows: Randomly sample equal numbers of source domain images and target domain images to construct a balanced dataset; the number of samples is set to the total number of samples in the target domain dataset, that is, all target domain data is added to the balanced dataset; The transfer model uses a cosine similarity-based classifier and adds an additional contrastive representation embedding branch; The weight matrix of the classifier It is expressed as [ω1,ω2,…,ω c ],in is the class weight vector for each category; the output of the classifier is the similarity score P between the input feature x and each class weight vector: Among them, p i,j is the similarity score between the i-th candidate region of x and the weight vector of category j, and α represents the scaling factor; The contrastive representation embedding branch adopts a multi-layer perceptron structure to encode ROI features into contrastive representation embedding vectors. By measuring the similarity scores between different candidate regions, a loss function is designed to make samples of the same category more compact in the feature space, while samples of different categories gradually move away from each other in the feature space: where z i is the contrast representation embedding vector of the i-th ROI feature encoding, u i Represents the intersection over union score IoU,y with the matching bounding box label i Represents the category label; Yes and y i The number of candidate regions of the same category, τ is the scaling factor; represents the normalized eigenvector, Measures the cosine similarity between the i-th and j-th candidate regions; f(u i ) to filter the candidate areas: where g(·) assigns different weight coefficients to objects with different IoU scores; φ is the IoU threshold; The joint loss function in the fine-tuning stage is expressed as: where λ is The weight factor of The migration model is initialized using the basic model parameters obtained in step 2. During fine-tuning, the parameters of the feature extraction network are fixed, and the parameters of the remaining modules are optimized through iterative training of the gradient descent algorithm until the model accuracy no longer improves or meets the requirements.

5. The method for locating and identifying unexploded bomb craters driven by transfer learning according to claim 4, characterized in that: The φ is set to 0.7, and the λ is set to 0.

5.

6. The method for locating and identifying unexploded bomb craters driven by transfer learning according to claim 4, characterized in that: The step 4 is specifically as follows: Given a set of test images, input the images into the model trained in step 3 to obtain a set of prediction results {(x1 i ,y1 i ,x2 i ,x2 i ,c i ,p i )|i=1,…,M}, where (x1 i ,y1 i ,x2 i ,x2 i ) is the bounding box coordinate of the i-th detected object, c i and p i Represent the corresponding prediction category and confidence score respectively; The confidence threshold is set to δ. If the confidence score of a target is higher than the threshold, the target is considered to be credible, that is, there may be unexploded bombs.

Images