Small sample explosion accident key broken equipment fragment identification method, system and equipment

Abstract

The application discloses a small sample explosion accident key damaged equipment fragment identification method, system and equipment, first, collecting and marking actual explosion scene fragment image data, fine tuning training is carried out to the base model, and a fragment identification model is obtained; then, according to the fragment identification model, the unmarked data in the current explosion scene is inferred, and the fragment identification result is counted and verified; if the verification result meets the identification accuracy, the fragment identification model is used to identify the fragments in the explosion scene; if the verification result does not meet the identification accuracy, the new data after inference is processed and added to the small sample data set, and the process of step 1 is repeated by using the new small sample data set to carry out new fine tuning optimization training on the fragment identification model. The application adopts the combination mode of meta learning and transfer learning, and is based on the base model constructed by rich samples, so that a small amount of explosion fragment samples are used to realize intelligent identification of fragments of key damaged equipment in different explosion scenes.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology and relates to a fragment recognition method, system and device, particularly to an intelligent fragment recognition method, system and device for key equipment at explosion accident sites based on small sample data. Background Technology

[0002] Explosions often cause significant loss of life and property, requiring not only proactive prevention and control but also thorough investigation and analysis of the causes to provide data for decision-making in further improving accident prevention measures. Source tracing analysis at explosion sites typically requires combining information on the damage to key equipment, such as fragments and explosion traces. This often necessitates substantial manpower and resources to search for and identify fragments of critical equipment. Furthermore, due to the inherent danger and complexity of explosion sites, searching for fragments of critical tank equipment is particularly challenging. Therefore, an advanced and rapid intelligent fragment identification method is needed to provide fragment identification results for post-explosion accident scenarios, thereby improving the efficiency of accident source tracing analysis.

[0003] With the rapid development of artificial intelligence technology, intelligent object recognition methods have emerged in an endless stream, becoming a hot topic in modern research. In recent years, the use of convolutional neural networks for image data classification, recognition, and segmentation has achieved very effective results and has gradually been applied and developed in daily life. However, deep learning models often require a large amount of labeled datasets as support, while real-world visual data exhibits a significant long-tail effect, with the most data-rich categories accounting for the majority of all categories. In certain specific application scenarios, some scarce categories may be difficult to acquire and label due to factors such as privacy, security, and high labeling costs. Examples include military remote sensing detection, disease diagnosis, and defective product detection in industrial production. This poses a challenge to the further development of the field of computer vision. At the same time, the sample size of debris datasets in explosion scenarios is small, and the sample data varies across different explosion scenarios. This makes the generalization ability of deep learning models weak, meaning lower recognition accuracy, and it is difficult to meet the requirements of complex explosion scenarios. Therefore, there is a need for a technology that can quickly transfer recognition tasks across different explosion scenarios using only a small number of samples, ensuring debris recognition accuracy while adapting to different debris types in different explosion scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying critical equipment fragments at explosion accident sites based on small samples. This method addresses the problem of difficulty in fragment identification when there are few critical equipment fragment samples in different explosion scenarios, reduces the dependence of deep learning models on large-scale data, and improves the generalizability of fragment identification models in different scenarios.

[0005] The technical solution adopted by the present invention is: a method for identifying fragments of critical damaged equipment in a small-sample explosion accident, comprising the following steps:

[0006] Step 1: Collect and label actual explosion scene debris image data, fine-tune the base model, and obtain a debris recognition model;

[0007] Step 2: Based on the debris recognition model obtained in Step 1, infer the unlabeled data in the current explosion scene and perform statistical verification on the debris recognition results;

[0008] If the verification result meets the recognition accuracy, the debris recognition model is used to identify debris in the explosion scene;

[0009] If the verification result does not meet the recognition accuracy, the inferred new data will be processed and added to the small sample dataset. Then, the process of step 1 will be repeated using the new small sample dataset to fine-tune and optimize the fragment recognition model.

[0010] Preferably, the base model in step 1 includes a feature extraction module, a Neck module, and a Head module;

[0011] The feature extraction module uses a twin backbone network to extract features from the query branch and the support branch respectively. The twin backbone network is trained with shared weights.

[0012] The Neck module adopts a C2f structure, which ensures the model is lightweight while constructing richer gradient information;

[0013] The Head module, in a decoupled manner, separates the bounding box position and target category regression into two branches;

[0014] The query branch output is processed by the Neck and Head modules to obtain the encoding result; the support branch output uses RoIAlign pooling to obtain the support feature vector result; finally, the support feature vector is fused with the query branch encoding result through a feature aggregation mechanism to detect the target location of the query image and the category discrimination result of the bounding box.

[0015] Preferably, the base model mentioned in step 1 is a trained model;

[0016] The specific training process includes the following steps:

[0017] (1) Construct training and testing datasets; the training and testing datasets contain image data and corresponding bounding box label data of the images;

[0018] (2) Train the base model using the training dataset;

[0019] The system utilizes a Siamese backbone network to extract features from both the query and support branches. The support image is processed by RoIAlign pooling to obtain support feature results, while the query image features are processed by the Neck part composed of C2f modules and the decoupling head to obtain the encoding results. The support feature results of the support image are fused with the encoding results through a feature aggregation mechanism to detect the target location of the query image and obtain the bounding box classification results.

[0020] With E base The training rounds are for the base model. The training process uses the SGD optimizer and trains with weight decay to obtain the corresponding base model. During the training process, the model training loss is calculated based on the obtained bounding box results and class results, which includes: classification loss BCEWithLogitsLoss, meta loss MetaLoss and boundary loss DistributeFocal Loss.

[0021] (3) Use the base model test dataset to fine-tune the training test of the base model, and adjust the base model trained in (2) according to the test results.

[0022] Preferably, in step (3), the base model trained in (2) is adjusted based on the test results. The test dataset is constructed using a few samples, and multiple random sampling of sample data is used to construct the query set and support set, which satisfy the condition of disjointness. Subsequently, the base model is subjected to transfer fine-tuning training, using fine-tuning training rounds E. finetune Fine-tuning training is performed, and the results of the fine-tuned model are validated. The average precision of these multiple random results is used as the standard. If the precision reaches the threshold λ, the performance of the base model meets the requirements for transferability.

[0023] Preferably, in step 1, actual explosion scene fragment image data is collected and labeled, a support set and a query set are constructed using the C-way-N-shot method, the base model is fine-tuned, the feature extraction backbone part in the base model is frozen, and the fine-tuning training epoch E is used. finetune The model is trained to obtain a fragmentation recognition model for identifying explosion debris.

[0024] As a preferred option, in step 2, inference is made on the unlabeled data in the current explosion scene. In the inference process, there is no need for a support set branch, and the image data is directly input into the query branch for inference.

[0025] The technical solution adopted by the system of the present invention is: a fragment identification system for critical damaged equipment in a small sample explosion accident, characterized in that it includes the following modules:

[0026] The base model fine-tuning training module allows users to collect and label actual explosion scene debris image data, fine-tune the base model, and obtain a debris recognition model.

[0027] The debris recognition module is used to fine-tune the debris recognition model obtained by the training module based on the base model, infer unlabeled data in the current explosion scene, and perform statistical verification on the debris recognition results.

[0028] If the verification result meets the recognition accuracy, the debris recognition model is used to identify debris in the explosion scene;

[0029] If the verification result does not meet the recognition accuracy, the inferred new data will be processed and added to the small sample dataset. The base model fine-tuning training module will then be executed using the new small sample dataset to perform new fine-tuning and optimization training on the fragment recognition model.

[0030] The technical solution adopted by the device of the present invention is: a fragment identification device for key damaged equipment in a small sample explosion accident, characterized in that it includes:

[0031] One or more processors;

[0032] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the method for identifying fragments of critical damaged equipment in a small-sample explosion accident.

[0033] This invention utilizes deep learning methods to address the issue of limited debris in explosion scenarios. Based on existing rich datasets, it proposes a method for identifying debris from critical equipment at explosion sites that combines transfer learning with meta-learning. This method can quickly transfer the target domain across various explosion scenarios, identify debris data from critical equipment, and continuously optimize and improve the identification accuracy by introducing an incremental optimization strategy, thus providing an intelligent debris identification means for explosion site tracing and analysis. Attached Figure Description

[0034] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0035] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0036] Figure 2 This is a flowchart of the base model training process according to an embodiment of the present invention. Detailed Implementation

[0037] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0038] This embodiment primarily utilizes deep learning and, considering the search for debris during the tracing of key equipment in explosion scenarios, proposes a debris identification method, system, and device based on few samples. This embodiment leverages abundant samples to train the model's feature extraction capabilities, then combines transfer learning and meta-learning to achieve transfer across different target domains and few-sample identification. Furthermore, it provides a model re-optimization mode to further improve the accuracy of few-sample identification.

[0039] Please see Figure 1 This embodiment provides a method for identifying fragments of critical damaged equipment in a small-sample explosion accident (taking circular fragment data as an example), which includes the following steps:

[0040] Step 1: Construct a base model training dataset using a public object detection dataset containing a wide variety of types and quantities. The training dataset includes image data and corresponding bounding box labels for the images. The dataset is divided according to different categories; one category constitutes the base model training set, while the other category data forms a small sample dataset to constitute the base model test set.

[0041] In one implementation, a training set and a test set can be constructed using a public dataset according to categories, and a support set and a query set can be constructed simultaneously for the training set and the test set. The support set is obtained by cropping images from the dataset using bounding boxes to obtain data samples for each category, used to train the model for feature extraction of the target. The query set is used to verify the model's target and validate the extraction performance. The two sets are proportionally allocated and do not overlap. The support set and query set in the training set are constructed with rich samples, while the test set is constructed with fewer samples.

[0042] In one implementation, the publicly available dataset VOC is used as a specific example. It contains rich sample images and labeled data across 20 different categories. Fifteen categories are randomly selected as the base training set, and five as the test set. In this implementation, aeroplane, bicycle, boat, bottle, car, cat, chair, dining table, dog, motorbike, person, potted plant, sofa, train, and TV monitor are used as the base model training categories, while bird, bus, cow, horse, and sheep are used as the base model test categories. For the base model training set, the dataset is divided into a base query set and a base support set. The base support set is constructed solely from image data cropped from the bounding boxes, using the results of each class's small images. This support set is then combined with the bounding box data to form the base support set. The query set does not require this operation; therefore, each query can contain multiple target data. Both the training and validation sets of this public dataset are processed in the same way. The number of query datasets and support sets for each category in the base model training and validation sets are shown in Table 1 below. The image data paths, bounding box positions, and type information are stored in a txt file.

[0043] Table 1

[0044]

[0045]

[0046] Step 2: Process the training dataset of the base model in Step 1, and train the base model using a few-shot training framework based on meta-learning.

[0047] In one implementation method, the specific implementation is as follows:

[0048] Pre-training is performed based on the processed base dataset, using a C-class N-sample approach (where N is typically the number of small samples), where C is the number of classes in the base training set and N is the number of support sets. First, the training data reading section includes data processing and data augmentation. Data augmentation is only randomly enabled on the training set, while the validation data is not augmented. In the data processing section, both the input image data and the bounding box data are processed simultaneously. The image data is normalized using the mean and variance of natural images and scaled to a uniform input size [w]. input H input The marker boxes are also normalized and marked with [x]. center y center Composed of the form of [x, w, h, class], where [x center y center[ ] represents the center of the bounding box, w and h are the width and height of the bounding box, and class is the category detected by the bounding box. Regarding data augmentation, according to a certain random ratio r... transform (Indicating the data augmentation ratio) Image transformation operations are performed on a portion of the data to obtain new data, including: vertical flipping, horizontal flipping, and color gamut conversion. After completing the image transformation and augmentation operations, mosaic and mixup operations (two data augmentation methods: mosaic enhancement and multi-image mixing enhancement) are applied, respectively according to the ratio r. mosaic r mixup (Indicates the ratio of the two data augmentation methods) Operates on both image data and bounding boxes simultaneously.

[0049] In one implementation, the YOLOv8 framework is used as the training framework, such as... Figure 2 As shown, its structure includes a feature extraction module, a Neck module, and a Head module. In the feature extraction module, a Siamese backbone network is used to extract features from both the query branch and the support branch, with the Siamese backbone network trained using shared weights. The Neck part adopts a C2f structure, ensuring a lightweight model while constructing richer gradient information. The Head part decouples the bounding box position and target category regression into two branches. The query branch output is processed by the Neck and Head modules to obtain the encoded result; the support branch output uses RoIAlign pooling to obtain the support feature vector result; finally, a feature aggregation mechanism is used to fuse the support feature vector with the query branch encoding result to detect the target location in the query image and the bounding box category determination result.

[0050] The model loss is calculated based on the obtained bounding box and category results. It includes: classification loss BCEWithLogitsLoss, meta loss MetaLoss, and boundary loss Distribute Focal Loss. The formula for calculating the loss function is shown below:

[0051]

[0052] For each of the C categories, x i There is a binary tag y i ∈0,1, the predicted probability of each sample is p i σ(x) is the Sigmoid function log is the natural logarithm.

[0053] The bounding box loss uses the Distributed Focal Loss regression loss:

[0054]

[0055]

[0056] Meanwhile, for the extracted support vectors, to encourage support image features to belong to the corresponding target category, cross-entropy loss is used to implement the meta-learning loss MetaLoss:

[0057]

[0058] Where N is the number of samples, y i For the truth value category, x i For the output category.

[0059] The base model is trained using E base In each training round, the SGD optimizer is used to train the model using weight decay, resulting in the corresponding basic data model.

[0060] In one implementation, the query set and support set of the base model training set constructed in step 1 are used as an example, where C is 15 and N is 5. Therefore, during the training phase, each query branch corresponds to 5 support branch data for training, and the data is input in a randomly sampled manner during the training process. The input size of the query set image [w] input H input The [640, 640] set is used, while the support set [w] is supported. support H support The feature is set to [320, 320]. Query set data augmentation ratio r mosaic r mixup r transfom The values ​​are 0.5, 0.5, and 0.5 respectively.

[0061] Training Round E base A 200-fold learning curve was used. The SGD optimizer was employed during training, with weight decay using cosine similarity. The initial learning rate was set to 0.01, and it decayed every 20 training epochs with a decay parameter of 0.0005. During training, the model was saved and validated every 10 epochs. On the validation set, mAP (Mean Average Precision), F1 score, Precision, and Recall were calculated to assess the effectiveness of the base model. Training log files were saved to record the changes in loss during training. The formulas for these metrics are as follows:

[0062]

[0063]

[0064]

[0065] In this model, TP represents the number of correctly detected positive fragment samples, FP is the number of samples that should have been fragments but were falsely detected as negative samples, and FN is the number of samples that were detected as fragments but whose true type is not fragments. AP represents the average value of the detector across all recall conditions, corresponding to the area under the Precision-Recall curve, while mAP is the average AP over categories, which can be used to evaluate multi-type recognition models.

[0066] During the training of the base model, this example enhances feature extraction capabilities and accelerates training speed by loading pre-trained weights from the backbone network. The model with the best performance during training is ultimately selected as the final base model. In this embodiment, the optimal base model achieves mAP, F1, Precision, and Recall metrics of 94.2%, 0.91%, 96.2%, and 86.3%, respectively.

[0067] Step 3: Use the base test set data to fine-tune the training test of the base model, and adjust the base model trained in Step 2 according to the test results.

[0068] The base model's transfer learning capability is tested using test categories from the base dataset that were not used in training as fine-tuning training samples. The base test set is constructed with a small sample size, using randomly selected data to build the query set and support set, which also satisfy the disjointness condition. The base model is then fine-tuned using a limited number of Efinetune training epochs, and the results are validated. Multiple random fine-tuning training sessions are performed, using the average precision of all sampled model tests as the validation metric. If the precision reaches a threshold λ, the base model performance meets the transferability requirement, and step 4 is performed; otherwise, the parameters are adjusted, and step 2 is repeated.

[0069] In one implementation, the remaining 5 classes of samples not involved in the base model training in step 1 are used as test classes, with N = 5. Therefore, a 5-shot dataset is constructed, with each query set corresponding to 5 support sets. In this embodiment, the base model test class dataset is randomly selected. The query set randomly selects 5 samples from each class, the support set selects 25 small image samples, and the forward inference test set selects 5 samples (used to verify model accuracy). These three datasets are disjoint and different. Ten random samplings are performed, and the average accuracy of these 10 samplings is used to evaluate the model's transfer performance, E. finetune With a learning rate of 0.01 and a learning value of 20, the average accuracy of the final forward inference test samples is 95.1%, which meets the accuracy requirements.

[0070] Step 4: Process the actual explosion scene debris data. After being labeled by professionals, determine the types of debris that need to be detected in the scene. If the dataset size meets the requirements, construct the support set and query set in a C-way-N-shot manner. If the conditions are not met, use data augmentation to increase the dataset size.

[0071] First, a small amount of debris image data from the explosion site is collected and labeled. Experts then determine the types of debris requiring identification. If the dataset is too small, data augmentation operations such as flipping and color gamut conversion are used to increase the dataset size to a threshold. The target domain support set and query set are then constructed using a C-way-N-shot approach.

[0072] Fine-tuning training is performed using the base model obtained after testing in step 3. Fine-tuning is typically faster and more accurate, thus eliminating the need for a large and rich dataset like base model training. By simply freezing certain modules, the model can be quickly transferred to different scenarios. In this invention, the feature extraction module is frozen during training, while other modules remain unchanged. Typically, some layers of the base model are frozen, and their parameters are not updated during fine-tuning to preserve the knowledge learned from the original dataset. This allows for fewer fine-tuning training epochs (E). finetune Fine-tuning training is performed, which involves readjusting hyperparameters such as the learning rate and batch size to improve the model's performance and convergence speed. The final result is an explosion fragment recognition model for the current target domain, which is then validated using test data. The detection accuracy is calculated statistically. If the accuracy does not reach the threshold λ, the model is fine-tuned again by changing the parameters until the model reaches the required performance. Here, the model performance is mainly tested using the Precision metric.

[0073] In one implementation, the parameters of the query set and support set constructed in step 1 are set as follows: C = 15 and N = 5. In this embodiment, taking 26 collected images of circular debris from the explosion site as an example, the required debris type is Class 1, the number of debris target boxes is 73, and the number of images meets the conditions for fine-tuning training. Therefore, a target domain support set and a query set are constructed, with 8 images in the query set and 38 images in the support set after processing. This embodiment of the invention uses training rounds E. finetune The relevant transfer model was obtained with parameters of 30 and a learning rate of 0.01, and the model test dataset consisted of 5 images. After training, the model was tested using the test set, and the precision was 95.3%, which meets the requirement of over 95% accuracy in actual scene recognition. Therefore, this model can be used for intelligent identification of scene debris.

[0074] Step 5: Use the small sample model of the current target domain to infer the unlabeled data in the scene, perform statistical verification on the results after the data is labeled, process the inferred data, and after reaching a certain number, add it to the small sample dataset to fine-tune and optimize the model, forming an incremental optimization strategy.

[0075] Based on the fragment recognition model obtained in step 4, inference is performed on the image data in the scene. At this point, support set branching is not required; test data is directly input into the query branch for inference and to obtain the recognition result. The recognition result is displayed as a combination of bounding boxes and confidence scores. The fragment recognition accuracy is then statistically analyzed. After processing, the test data is added to a small sample dataset, accumulating a certain number of Num... new Then, repeat the fine-tuning and optimization training process in step 4. If the new model meets the threshold requirement, replace the previous model; otherwise, adjust the parameters and continue optimization. If the number of new data points does not reach Num... new If the current fragmentation identification model is not found, the current model will continue to be used until the amount of new data reaches the required level.

[0076] In one implementation, as the circular fragment data is continuously identified, new sample data accumulates. After manual review by professionals, this data is added to the target domain small sample dataset from step 4. When the number Num new After reaching 30 iterations, a new small sample dataset for the current target domain is reconstructed. The new small sample dataset contains 16 support images and 71 query images. After training for 20 rounds with a learning rate of 0.001, a new circular fragment recognition model is obtained, with a recognition accuracy of 96.2%. Therefore, this model is used to replace the previous model as the model for real-time fragment recognition.

[0077] Step 6: Use the few-sample fragment recognition model to identify fragments in the scene. If new types need to be added during the recognition process, rebuild the few-sample dataset and fine-tune the model training, repeating the training process of Step 4.

[0078] The current target domain model is used to identify fragments in the scene. Similarly, if new types of image data are available, step 4 is repeated to process the new category data and combine it with the previous data to create a new small sample dataset for testing. If no new categories need to be added, the current model is used, and step 5 determines whether a new dataset needs to be built and a new model trained.

[0079] This embodiment also provides a fragment identification system for critical damaged equipment in a small-sample explosion accident, including the following modules:

[0080] The base model fine-tuning training module allows users to collect and label actual explosion scene debris image data, fine-tune the base model, and obtain a debris recognition model.

[0081] The debris recognition module is used to fine-tune the debris recognition model obtained by the training module based on the base model, infer unlabeled data in the current explosion scene, and perform statistical verification on the debris recognition results.

[0082] If the verification result meets the recognition accuracy, the debris recognition model is used to identify debris in the explosion scene;

[0083] If the verification result does not meet the recognition accuracy, the inferred new data will be processed and added to the small sample dataset. The base model fine-tuning training module will then be executed using the new small sample dataset to perform new fine-tuning and optimization training on the fragment recognition model.

[0084] This embodiment also provides a device for identifying fragments of critical damaged equipment in a small-sample explosion accident, including:

[0085] One or more processors;

[0086] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the method for identifying fragments of critical damaged equipment in a small-sample explosion accident.

[0087] In this embodiment, circular fragment data is used as the primary recognition type, and large can fragments are added as a new recognition type during the later recognition process. After being labeled by professionals, a small sample dataset for this new type is created. Due to the small size of this dataset, it is expanded to obtain a query set of 7 images and a support set of 33 images. Subsequently, it is merged with the circular fragment dataset to construct a new small sample dataset, and the fine-tuning training process in step 4 is repeated using this dataset as the training dataset to obtain recognition models for these two types of fragments. Statistical analysis shows that the recognition accuracy for circular fragments is 96.1%, and the recognition accuracy for large can fragments is 95.4%. This meets the fragment recognition accuracy requirements for the current scene, therefore, this model is used as the recognition model for the current scene.

[0088] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0089] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for identifying fragments of critical damaged equipment in a small-sample explosion accident, characterized in that, Includes the following steps: Step 1: Collect and label actual explosion scene debris image data, fine-tune the base model, and obtain a debris recognition model; The base model includes: a feature extraction module, a Neck module, and a Head module; The feature extraction module uses a twin backbone network to extract features from the query branch and the support branch respectively. The twin backbone network is trained with shared weights. The Neck module adopts a C2f structure, which ensures the model is lightweight while constructing richer gradient information; The Head module, in a decoupled manner, separates the bounding box position and target category regression into two branches; The output of the query branch is processed by the Neck and Head modules to obtain the encoded result; the output of the support branch is obtained by RoIAlign pooling operation to obtain the support feature vector result. Finally, the supporting feature vectors and query branch encoding results are fused through a feature aggregation mechanism to detect the target location of the query image and the category discrimination result of the bounding box; Step 2: Based on the debris recognition model obtained in Step 1, infer the unlabeled data in the current explosion scene and perform statistical verification on the debris recognition results; If the verification result meets the recognition accuracy, the debris recognition model is used to identify debris in the explosion scene; If the verification result does not meet the recognition accuracy, the inferred new data will be processed and added to the small sample dataset. Then, the process of step 1 will be repeated using the new small sample dataset to fine-tune and optimize the fragment recognition model.

2. The method for identifying fragments of critical damaged equipment in a small-sample explosion accident according to claim 1, characterized in that: The base model mentioned in step 1 is a trained model; The specific training process includes the following steps: (1) Construct training datasets and test datasets; the training datasets and test datasets contain image data and corresponding bounding box label data of the images; (2) Train the base model using the training dataset; The system utilizes a Siamese backbone network to extract features from both the query and support branches. The support image is processed by RoIAlign pooling to obtain support feature results, while the query image features are processed by the Neck part composed of C2f modules and the decoupling head to obtain the encoding results. The support feature results of the support image are fused with the encoding results through a feature aggregation mechanism to detect the target location of the query image and obtain the bounding box classification results. by The training rounds are for the base model. The training process uses the SGD optimizer and trains with weight decay to obtain the corresponding base model. During the training process, the model training loss is calculated based on the obtained bounding box results and class results, which includes: classification loss BCEWithLogitsLoss, meta loss MetaLoss, and boundary loss Distribute FocalLoss. (3) Use the base model test dataset to fine-tune the training test of the base model, and adjust the base model trained in (2) according to the test results.

3. The method for identifying fragments of critical damaged equipment in a small-sample explosion accident according to claim 2, characterized in that: In step (3), the base model trained in (2) is adjusted according to the test results. The test dataset is constructed in a few-sample manner, and the query set and support set are constructed by randomly sampling data multiple times. The query set and support set satisfy the condition of disjointness. Subsequently, the base model was subjected to transfer fine-tuning training, utilizing fine-tuning training epochs. Perform fine-tuning training and validate the results of the fine-tuned model; With the average precision of these multiple random results As a standard, if Reaching the threshold If the performance of the base model meets the requirements for transferability, then the performance of the base model is satisfactory.

4. The method for identifying fragments of critical damaged equipment in a small-sample explosion accident according to claim 1, characterized in that: In step 1, actual explosion scene debris image data is collected and labeled. A support set and query set are constructed using a C-class N-sample method. The base model is then fine-tuned. The feature extraction backbone part of the base model is frozen, and the training epochs are fine-tuned. Training is performed to obtain a fragmentation recognition model for identifying explosion fragments; where N is the number of small samples.

5. The method for identifying fragments of critical damaged equipment in a small-sample explosion accident according to any one of claims 1-4, characterized in that: In step 2, inferences are made about the unlabeled data in the current explosion scene. No support set branch is needed during the inference process; the image data is directly input into the query branch for inference.

6. A fragment identification system for critical damaged equipment in a small-sample explosion accident, characterized in that, Includes the following modules: The base model fine-tuning training module allows users to collect and label actual explosion scene debris image data, fine-tune the base model, and obtain a debris recognition model. The base model includes: a feature extraction module, a Neck module, and a Head module; The feature extraction module uses a twin backbone network to extract features from the query branch and the support branch respectively. The twin backbone network is trained with shared weights. The Neck module adopts a C2f structure, which ensures the model is lightweight while constructing richer gradient information; The Head module, in a decoupled manner, separates the bounding box position and target category regression into two branches; The output of the query branch is processed by the Neck and Head modules to obtain the encoded result; the output of the support branch is obtained by RoIAlign pooling operation to obtain the support feature vector result. Finally, the supporting feature vectors and query branch encoding results are fused through a feature aggregation mechanism to detect the target location of the query image and the category discrimination result of the bounding box; The debris recognition module is used to fine-tune the debris recognition model obtained by the training module based on the base model, infer unlabeled data in the current explosion scene, and perform statistical verification on the debris recognition results. If the verification result meets the recognition accuracy, the debris recognition model is used to identify debris in the explosion scene; If the verification result does not meet the recognition accuracy, the inferred new data will be processed and added to the small sample dataset. The base model fine-tuning training module will then be executed in a new fine-tuning and optimization training to improve the fragment recognition model.

7. A fragment identification device for critical damaged equipment in a small-sample explosion accident, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for identifying fragments of critical damaged equipment in a small sample explosion accident as described in any one of claims 1 to 5.

Images