Remote sensing image target detection method, system and device based on progressive self-training and medium thereof

Abstract

The application discloses a kind of based on gradual self-training remote sensing image target detection method, system, equipment and medium thereof, belong to computer vision field, including the following steps: training initial target detection model;Initial target detection model is used to generate boundary box level pseudo label for unmarked image data, and the boundary box level pseudo label is filtered using noise pseudo label filtering strategy, and the image data with reliable boundary box level label is obtained;According to initial training data set and the image data with reliable boundary box level label, reliable target detection model is obtained by model training;Reliable target detection model is used to continuously train and iterate optimization, until the image data with reliable boundary box level label quantity stops growth, and high-precision target detection model is obtained, and target detection is completed using high-precision target detection model.The application can solve the problem that the model detection precision is reduced when the prior art reduces the dependence on manually labeled images for model training.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and involves technologies such as pattern recognition, image processing and deep learning. Specifically, it relates to a remote sensing image target detection method, system, device and medium based on progressive self-training. Background Technology

[0002] In current remote sensing applications, target detection technology plays a crucial foundational role. It not only helps monitor the utilization of natural resources and the development of urban construction, but also plays a vital role in disaster monitoring and environmental protection. However, accurate target detection largely depends on large-scale training data with bounding box annotations. But the scale, shape, and occlusion of objects in remote sensing images are constantly changing, thus the complexity and diversity of objects in remote sensing images pose a significant challenge to manual bounding box annotation in this field.

[0003] In recent years, researchers have shown increasing interest in semi-supervised object detection (SSOD) as a potential solution to alleviate the burden of data annotation. SSOD utilizes a combination of finitely annotated datasets and large-scale unlabeled datasets to optimize model performance and reduce the model's dependence on labeled data. However, SSOD assumes that both labeled and unlabeled data are sampled from the same underlying distribution, an assumption that is not always valid in real-world scenarios. Furthermore, SSOD performance depends not only on the manual annotation of the data subset but also heavily on the quality of the labeled data. Therefore, to further alleviate the burden of data annotation, image-level weakly supervised object detection (WSOD) has received considerable attention. WSOD relies on image-level supervision to identify object regions, thereby eliminating the need for precise bounding box annotations. However, WSOD ignores the intrinsic properties captured by bounding box annotations, resulting in inaccurate learned detection knowledge. Its performance lags behind fully supervised object detection (FSOD), and it faces challenges in accurately locating objects. Moreover, the typical feature representations learned from weakly supervised settings only capture salient parts of representative objects, deviating from the expected detection results. Summary of the Invention

[0004] The purpose of this invention is to provide a remote sensing image target detection method, system, device and medium based on progressive self-training, which solves the problem of reduced model detection accuracy when reducing reliance on manually labeled images for model training in existing technologies.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A remote sensing image target detection method based on progressive self-training includes the following steps: S1: Collect the initial training dataset and train the model based on the initial training dataset, while minimizing the loss to obtain the initial object detection model; S2: Use the initial target detection model to generate bounding box-level pseudo-labels for the unlabeled image data, and use a noise pseudo-label filtering strategy to filter the bounding box-level pseudo-labels to obtain image data with reliable bounding box-level labels; S3: Train the model based on the initial training dataset and image data with reliable bounding box level labels, and minimize the loss to obtain a reliable object detection model; After replacing the initial target detection model with the reliable target detection model, S2~S3 are repeated sequentially until the number of image data with reliable bounding box level labels stops growing, thus obtaining a high-precision target detection model. The target detection is then performed using the high-precision target detection model.

[0006] Furthermore, in S1, the initial training dataset is collected through the following steps: Image patches are obtained by overlapping and cropping publicly available remote sensing image datasets; After dividing the image blocks into multiple categories using an unsupervised clustering algorithm, clustered image data is obtained; Delete image blocks that do not belong to the target category from the clustered image data to obtain the target image block; An unsupervised instance detection model is used to generate bounding box-level pseudo-labels for the target image patch. Then, geometric constraints and the unsupervised clustering algorithm are applied sequentially to clean up the bounding box-level pseudo-labels of the target image patch, resulting in cleaned bounding box-level pseudo-labels and their corresponding image patches. The purified bounding box-level pseudo-labels and their corresponding image patches are used as the initial training dataset.

[0007] Furthermore, the step of dividing the image patch into multiple categories using an unsupervised clustering algorithm to obtain clustered image data specifically includes: First, the features of the image patch are extracted using the ResNet34 neural network. After the features are input into the K-means++ algorithm, the image patch is divided into multiple categories to obtain clustered image data. The number of the multiple categories is twice the number of image categories in the publicly available remote sensing image dataset.

[0008] Furthermore, the unsupervised clustering algorithm adopts the K-means++ algorithm, and the unsupervised instance detection model adopts the CutLER model; In S1 and S3, the model training is based on the YOLO v5x model.

[0009] Furthermore, the noise pseudo-label filtering strategy described in S2 specifically includes the following steps: The bounding box-level pseudo-labels are filtered using the geometric constraints, and then further filtered using a class-adaptive threshold filtering method to obtain bounding box-level pseudo-labels and their corresponding image data after removing class-related noise.

[0010] Furthermore, in S2, the noise pseudo-label filtering strategy further includes the following steps: Based on the bounding box-level pseudo-labels and their corresponding image data, the image category with the most image data is obtained through traversal. Then, the average confidence score of instances belonging to the image category with the most image data is calculated. Then, an average value is preset based on the average confidence score, and the preset average value is less than the average confidence score. Delete the bounding box-level pseudo-labels and their corresponding image data that are less than the preset average value.

[0011] Furthermore, in S2, the noise pseudo-label filtering strategy further includes the following steps: Based on the bounding box-level pseudo-labels and their corresponding image data, the image category with the most image data is obtained by traversal. A preset semantic correlation threshold is set between the image category with the most image data and other image categories; Image data in other image categories with a semantic relevance threshold greater than the threshold are identified as relevant targets, while image data in other image categories with a semantic relevance threshold less than the threshold are identified as irrelevant targets. The irrelevant targets are suppressed using constraint factors, and the average confidence score of the image category to which the relevant targets belong and the confidence score of the relevant targets are calculated. Delete relevant targets and their corresponding bounding box-level pseudo-labels that are lower than the average confidence score of the image category to which the relevant target belongs.

[0012] A remote sensing image target detection system based on progressive self-training, comprising: The initial object detection model building module is used to collect an initial training dataset, train the model based on the initial training dataset, and minimize the loss to obtain the initial object detection model. The noise pseudo-label filtering module is used to generate bounding box-level pseudo-labels for unlabeled image data using the initial target detection model, and to filter the bounding box-level pseudo-labels using a noise pseudo-label filtering strategy to obtain image data with reliable bounding box-level labels. A reliable object detection model building module is used to train the model based on the initial training dataset and image data with reliable bounding box level labels, while minimizing the loss to obtain a reliable object detection model. A progressive self-training module is used to replace the initial target detection model with the reliable target detection model for self-training until the number of image data with reliable bounding box level labels stops growing, thereby obtaining a high-precision target detection model, and using the high-precision target detection model to complete target detection.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor executes the computer program to implement the steps of the aforementioned method for detecting remote sensing images based on progressive self-training.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a remote sensing image target detection method based on progressive self-training.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a remote sensing image target detection method based on progressive self-training. First, an initial target detection model is trained using an initial training dataset. Then, the initial target detection model is used to generate bounding box-level pseudo-labels for unlabeled image data, and a noise pseudo-label filtering strategy is used to filter these pseudo-labels, resulting in image data with reliable bounding box-level labels. Based on the initial training dataset and the image data with reliable bounding box-level labels, a reliable target detection model is obtained. This reliable target detection model replaces the initial target detection model for self-training, resulting in a high-precision target detection model that performs target detection. The use of the initial target detection model to generate bounding box-level pseudo-labels and the processing with a noise pseudo-label filtering strategy enables remote sensing image target detection without the need for manual bounding box-level labeling, solving the time-consuming and labor-intensive operation of manual bounding box-level labeling. Furthermore, this invention continuously iterates the reliable target detection model until the amount of image data stops growing, ultimately obtaining a high-precision target detection model for target detection. Compared to existing weakly supervised methods, this high-precision target detection model exhibits better detection performance.

[0016] Furthermore, this invention uses the K-means++ algorithm to classify and obtain clustered image data. It has a fast convergence speed and high efficiency in processing large-scale remote sensing image data. At the same time, it has good clustering effect, which can accurately classify image data into different categories and reveal the inherent structure and pattern of the data.

[0017] Furthermore, this invention utilizes the unsupervised instance detection model CutLER to generate bounding box-level pseudo-labels for target image patches, which is compatible with detection models of different architectures and adaptable to various application scenarios and requirements; at the same time, it has the ability to identify multiple objects, accurately identify and distinguish different objects in the same scene, and improve the accuracy and efficiency of label generation.

[0018] Furthermore, this invention combines the YOLO v5x target detection model with remote sensing image target detection, which can improve detection accuracy and help to more accurately identify and locate targets in various scenarios; it enables this invention to process a large number of images in a short time, meeting the needs of real-time detection; and it can support the processing of high-resolution images, which is beneficial for capturing more detailed information. Attached Figure Description

[0019] Figure 1 This is a general framework diagram of the remote sensing image target detection method based on progressive self-training provided in this embodiment; Figure 2 The detection results are those of the remote sensing image target detection method based on progressive self-training provided in this embodiment. Figure 3 This is a schematic diagram of the structure of the remote sensing image target detection system based on progressive self-training provided in this embodiment; Figure 4 This is a structural diagram of the electronic device used in this invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0021] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this invention are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products, or devices.

[0022] This embodiment provides a remote sensing image target detection method based on progressive self-training, including the following steps: Step 1: Collect publicly available remote sensing image datasets, including RSOD (Remote Sensing Object Detection) dataset, LEVIR dataset, NWPU VHR-10.v2 dataset, DIOR (Object Detection in Optical Remote sensing images) dataset, UCAS-AOD (UCAS-High Resolution Aerial Object Detection) dataset, etc., and segment these images into image patches using an overlap cropping operation; The DIOR dataset in step 1 was not cropped because the target instances in the DIOR dataset are large, and overlapping cropping cannot guarantee the integrity of the targets.

[0023] Step 2: Apply the K-means++ unsupervised clustering algorithm to cluster the image patches obtained in Step 1 into 2N categories, where N represents the number of categories in the entire dataset; Specifically, in step 2, a ResNet34 neural network is first used to extract features from the image patches. The extracted features are then input into K-means++ to cluster the image patches into 2N classes. Although the cropped dataset has a simpler background, facilitating subsequent unsupervised image clustering and unsupervised instance detection, it inevitably generates additional categories and images with different backgrounds. For example, overlapping cropping of the RSOD dataset will produce targets such as airplanes, runways, grass, and buildings. Therefore, after inputting the extracted features into K-means++, the image patches are clustered into 2N classes instead of N classes.

[0024] Step 3: Delete image blocks in the clustered image data that do not include vehicle, oil tank, aircraft and ship categories. In this step, the main focus is on target detection of vehicles, oil tank, aircraft and ships, so clustered images of other categories are directly discarded.

[0025] Step 4: Use the unsupervised instance detection model CutLER to generate high-quality bounding box-level pseudo-labels for the remaining image patches from Step 3.

[0026] Step 5: Clean up the bounding box-level pseudo-labels using geometric constraints and unsupervised clustering algorithms. The cleaned bounding box-level pseudo-labels and their corresponding image patches constitute the initial training dataset InitData. Specifically, the geometric constraints are shown in equations (1), (2), and (3) below:

[0027] in, and This represents the coordinates of the center point of the rectangle. This indicates the width of the rectangle. Indicates the height of the rectangle. and These represent the width and height of the image, respectively. Considering the uniqueness of each data type, different constraints are applied to different data types. In this embodiment, the image blocks of vehicles, oil tanks, and ships are constrained by formulas (1), (2), and (3), respectively, with the parameters set as follows: =0.5, =0.25, and the aircraft image patch follows the constraints of formulas (1) and (2), with the parameters set as follows: =0.95, =0.5. After generating bounding box annotations, the K-means++ algorithm is applied to each extracted image class to filter out noisy pseudo-labels.

[0028] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0029] Step 6: Train the YOLO v5x detection model on the initial training dataset InitData and minimize the loss; Step 7: Use the trained YOLO v5x detection model to generate bounding box-level pseudo-labels for large-scale unlabeled image data; Step 8: Use the Noise Pseudo-Label Filter (NPLFilter) to filter out unreliable bounding box-level pseudo-labels and obtain image data with reliable bounding box-level labels; Specifically, the steps of the noise pseudo-label filtering strategy are as follows: Step 8.1: Use geometric constraints to filter out noise pseudo-labels, and use the above formula (1) for filtering; Step 8.2: After completing the noise pseudo-label filtering based on geometric constraints, the class-adaptive threshold filtering method is then applied to further refine the noise pseudo-labels. Specifically, the average confidence score of each class instance is calculated, and these average confidence scores are used as thresholds to filter out noise pseudo-labels related to each class. The class-adaptive threshold is expressed by the following formula (4): (4) in, Indicates belonging to class The average confidence score of all instances. Indicates belonging to class The total number of instances, It belongs to The class of The confidence scores for each instance. Note that the scores obtained are due to noise from the pseudo-labels. Higher than the true accuracy. Therefore, as shown in equation (4), for each category ,Will As a reference, a slightly lower threshold, i.e., the mean minus 0.2, was chosen as the final class-adaptive threshold. To filter out noise and fake labels.

[0030] Step 8.3: To further improve the quality of the remaining bounding box-level pseudo-labels, a strategy based on prior knowledge is adopted, specifically: First, calculate the number of instances of each category in the image, where the category with the highest frequency (e.g., the first...) is identified. The class is considered as the most likely scene to occur in the image to be detected; then, within the class of the most likely scene, the average confidence score of instances in the image data belonging to that class is calculated, and any value lower than this average confidence score is set as the average. Finally, any confidence score less than this average is considered as the first instance. All class instances will be discarded and their bounding box-level pseudo-labels will be removed.

[0031] Step 8.4: To filter out pseudo-labels that do not conform to the objective pattern, this invention uses a sample statistical method to obtain the semantic relevance between categories. The calculation formula is as follows: (5) in, Indicates that the target is included. A collection of images, It is a function for counting items. It is the set of all target categories. This represents the dataset obtained after inference on the previously trained model in the previous training round. The specific steps are: first, identify the category with the most instances in the image data; then, query the semantic relevance between this category and other categories; and finally... Set as a threshold for semantic association. Where greater than or equal to... The value of the target is considered a relevant target, less than The objective of the value is treated as an irrelevant objective, and finally, a constraint factor is introduced. Multiply by the attribute probability of the target to suppress irrelevant targets. After applying the constraints, the confidence score related to the target modification can be quantified as follows (6): (6) constraint factors in this embodiment Set to 0.8, the threshold for semantic association. Set to 0.05, any confidence score Less than ( Indicates the first The average confidence score of all instances of the class) The class instance will be discarded and its corresponding bounding box-level pseudo-label will be removed.

[0032] Step 9: Relearn the new YOLO v5x detection model using the combination of the initial training dataset InitData and image data with reliable bounding box level labels as new training data, and minimize the loss.

[0033] Step 10: Iterate through steps 7 through 9 until the number of training images stops growing.

[0034] Step 11: Use the image data with reliable bounding box level labels obtained in Step 10 and the initial training dataset InitData as training data to train the YOLO v5x detection model and complete high-precision object detection.

[0035] The methods provided in the above embodiments describe the datasets used and the experimental results: 1. Dataset (1) Evaluation Dataset: This embodiment verifies the effectiveness of the proposed remote sensing image target detection method on the NWPU VHR-10.v2 dataset. The NWPU VHR-10.v2 dataset contains 10 categories and a total of 650 images. 60 images containing vehicles, oil tanks, aircraft, and ships were randomly selected as test data. The remaining 590 images were used as unlabeled data.

[0036] (2) Unlabeled Datasets: Unlabeled images were obtained from various sources, including the training sets of RSOD, LEVIR, DOTA-v1.0, NWPU VHR-10.v2, DIOR, and UCAS-AOD. Although these images are labeled, they are treated as unlabeled data in this embodiment. The RSOD dataset consists of 976 images across four object categories: airplanes, oil tanks, playgrounds, and overpasses. LEVIR contains 22,000 images covering three object categories: airplanes, ships, and oil tanks. UCAS-AOD is divided into two categories: airplanes and vehicles. There are 1,000 airplane images containing 7,482 instances, and 510 vehicle images containing 7,114 instances. The DOTA-v1.0 dataset contains 2,806 images ranging in size from 800*800 to 400*4000. To achieve accurate detection, the image was cropped into 640*640 blocks with a stride of 128, resulting in a total of 60,096 images after cropping.

[0037] 2. Experimental Results Table 1 shows a performance comparison between the proposed progressive self-training-based remote sensing image target detection method and state-of-the-art weakly supervised target detection methods, demonstrating that the proposed method achieves better results than these weakly supervised methods. In Table 1, this embodiment is compared with state-of-the-art weakly supervised target detection methods on the NWPU VHR-10.v2 test set, where the performance parameters corresponding to Ours in the last row are the performance verification of the proposed method.

[0038] besides, Figure 2 The system also displays visualized detection results, demonstrating that the invention can accurately detect small targets.

[0039] Table 1. Performance comparison of the progressive self-training-based remote sensing image target detection method with other target detection methods. WSDDN 0.3008 0.4172 0.0360 0.3498 0.2760 OICR 0.1366 0.6735 0.0373 0.5716 0.3548 PCL 0.2600 0.6376 0.1567 0.0250 0.2766 MELM 0.8086 0.6930 0.0868 0.1048 0.4233 DCL 0.7270 0.7425 0.3500 0.3705 0.5475 TCANet 0.8943 0.7818 0.2830 0.7842 0.6858 DPLG 0.8090 0.7830 0.0870 0.1050 0.4460 SPG+OICR 0.4280 0.7256 0.2475 0.5882 0.4973 AE-IS 0.9100 0.8820 0.4580 0.7830 0.7583 Ours 0.9950 0.7430 0.7440 0.9240 0.8510 Based on the above, this invention proposes an initial training data pseudo-label generation strategy. This strategy automatically generates high-quality bounding box annotations for the initial training data using unsupervised image clustering, unsupervised instance detection, and geometric constraints. This significantly reduces the cost and complexity associated with data annotation, providing crucial support for subsequent iterative training phases. The generated annotations are used as training data for iterative learning. The detector and pseudo-labels are refined through progressive self-training and the integration of the Noise Pseudo-Label Filter (NPLFilter) constructed in this invention, thereby achieving accurate target detection in remote sensing images. The NPLFilter cleanses the pseudo-labels by integrating geometric constraints, prior knowledge, and class-adaptive thresholds. Without using any image-level labels or bounding box-level annotations of the target data, the remote sensing image target detection method proposed in this invention achieves better detection performance than existing weakly supervised remote sensing target detection methods.

[0040] This embodiment also provides a remote sensing image target detection system based on progressive self-training, including an initial target detection model establishment module, a noise pseudo-label filtering module, a reliable target detection model establishment module, and a progressive self-training module; Specifically, the initial object detection model building module collects an initial training dataset and trains the model based on the initial training dataset, while minimizing the loss to obtain the initial object detection model; the noise pseudo-label filtering module uses the initial object detection model to generate bounding box-level pseudo-labels for unlabeled image data, and uses a noise pseudo-label filtering strategy to filter the bounding box-level pseudo-labels to obtain image data with reliable bounding box-level labels; the reliable object detection model building module trains the model based on the initial training dataset and the image data with reliable bounding box-level labels, while minimizing the loss to obtain the reliable object detection model; the progressive self-training module uses the reliable object detection model to replace the initial object detection model for self-training until the number of image data with reliable bounding box-level labels stops growing, thus obtaining a high-precision object detection model, which is then used to complete object detection.

[0041] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a remote sensing image target detection method based on progressive self-training.

[0042] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the remote sensing image target detection method based on progressive self-training in the above embodiments.

[0043] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A remote sensing image target detection method based on progressive self-training, characterized in that, Includes the following steps: S1: Collect the initial training dataset and train the model based on the initial training dataset, while minimizing the loss to obtain the initial object detection model; S2: Use the initial target detection model to generate bounding box-level pseudo-labels for the unlabeled image data, and use a noise pseudo-label filtering strategy to filter the bounding box-level pseudo-labels to obtain image data with reliable bounding box-level labels; S3: Train the model based on the initial training dataset and image data with reliable bounding box level labels, while minimizing the loss to obtain a reliable object detection model; After replacing the initial target detection model with the reliable target detection model, S2~S3 are repeated sequentially until the number of image data with reliable bounding box level labels stops growing, and a high-precision target detection model is obtained. The target detection is then completed using the high-precision target detection model. In S1, the initial training dataset is collected through the following steps: Image patches are obtained by overlapping and cropping publicly available remote sensing image datasets; After dividing the image blocks into multiple categories using an unsupervised clustering algorithm, clustered image data is obtained; Delete image blocks that do not belong to the detected target category from the clustered image data to obtain the target image block; An unsupervised instance detection model is used to generate bounding box-level pseudo-labels for the target image patch. Then, geometric constraints and the unsupervised clustering algorithm are applied sequentially to clean up the bounding box-level pseudo-labels of the target image patch, resulting in cleaned bounding box-level pseudo-labels and their corresponding image patches. The purified bounding box-level pseudo-labels and their corresponding image patches are used as the initial training dataset; The noise pseudo-label filtering strategy described in S2 specifically includes the following steps: The bounding box-level pseudo-labels are filtered using the geometric constraints, and then further filtered using a class-adaptive threshold filtering method to obtain bounding box-level pseudo-labels and their corresponding image data after removing class-related noise. In S2, the noise pseudo-label filtering strategy further includes the following steps: Based on the bounding box-level pseudo-labels and their corresponding image data, the image category with the most image data is obtained through traversal. Then, the average confidence score of instances belonging to the image category with the most image data is calculated. Then, an average value is preset based on the average confidence score, and the preset average value is less than the average confidence score. Delete the bounding box-level pseudo-labels and their corresponding image data that are less than the preset average value.

2. The remote sensing image target detection method based on progressive self-training according to claim 1, characterized in that, The step of dividing the image patch into multiple categories using an unsupervised clustering algorithm to obtain clustered image data specifically includes: First, the features of the image patch are extracted using the ResNet34 neural network. After the features are input into the K-means++ algorithm, the image patch is divided into multiple categories to obtain clustered image data. The number of the multiple categories is twice the number of image categories in the publicly available remote sensing image dataset.

3. The remote sensing image target detection method based on progressive self-training according to claim 1, characterized in that, The unsupervised clustering algorithm uses the K-means++ algorithm, and the unsupervised instance detection model uses the CutLER model. In S1 and S3, the model training is based on the Yolov5x model.

4. The remote sensing image target detection method based on progressive self-training according to claim 1, characterized in that, In S2, the noise pseudo-label filtering strategy further includes the following steps: Based on the bounding box-level pseudo-labels and their corresponding image data, the image category with the most image data is obtained by traversal. A preset semantic correlation threshold is set between the image category with the most image data and other image categories; Image data in other image categories with a semantic relevance threshold greater than the threshold are identified as relevant targets, while image data in other image categories with a semantic relevance threshold less than the threshold are identified as irrelevant targets. The irrelevant targets are suppressed using constraint factors, and the average confidence score of the image category to which the relevant targets belong and the confidence score of the relevant targets are calculated. Delete relevant targets and their corresponding bounding box-level pseudo-labels that are lower than the average confidence score of the image category to which the relevant target belongs.

5. A remote sensing image target detection system based on progressive self-training, characterized in that, A remote sensing image target detection method based on progressive self-training as described in any one of claims 1 to 4, comprising: The initial object detection model building module is used to collect an initial training dataset and train the model based on the initial training dataset, while minimizing the loss to obtain the initial object detection model; The noise pseudo-label filtering module is used to generate bounding box-level pseudo-labels for unlabeled image data using the initial target detection model, and to filter the bounding box-level pseudo-labels using a noise pseudo-label filtering strategy to obtain image data with reliable bounding box-level labels. A reliable object detection model building module is used to train the model based on the initial training dataset and image data with reliable bounding box level labels, while minimizing the loss to obtain a reliable object detection model. A progressive self-training module is used to replace the initial target detection model with the reliable target detection model for self-training until the number of image data with reliable bounding box level labels stops growing, thereby obtaining a high-precision target detection model, and using the high-precision target detection model to complete target detection.

6. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable in the processor. When the processor executes the computer program, it implements the steps of the remote sensing image target detection method based on progressive self-training as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the remote sensing image target detection method based on progressive self-training as described in any one of claims 1 to 4.

Images