Synthetic aperture radar image target detection method, device and equipment and storage medium

Abstract

A synthetic aperture radar image target detection method, device, equipment and storage medium are disclosed. The method comprises: acquiring a SAR image to be detected; using a denoising network to perform denoising processing on the SAR image; establishing an integrated target detection model, including the denoising network and a subsequent open-vocabulary target detection model, training the integrated model using an improved bounding box regression loss function constructed based on corner distance and side overlap information; inputting the denoised SAR image and text prompts into the trained open-vocabulary target detection model for feature extraction and target detection, and outputting a detection result. The present disclosure combines denoising technology and open-vocabulary target detection, solves the problem that spot noise is not fully processed in the current SAR target detection field, overcomes the limitation of traditional target detection on fixed categories, improves target detection accuracy, and expands the applicability and robustness of the model.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of synthetic aperture radar and machine vision, and particularly relates to a synthetic aperture radar image target detection method and device, equipment and storage medium. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) has important applications in military, geological and other fields due to its high resolution and all-weather, all-day working ability. SAR technology can work under various lighting and weather conditions, reflect the physical properties and surface roughness of objects, and provide different information from optical and infrared images. These advantages make SAR image target detection an increasingly important research field.

[0003] However, SAR image target detection faces unique challenges. Due to the coherence of SAR systems, SAR images inherently have speckle noise, which severely limits the effectiveness of detection algorithms. Compared with optical images, the features of targets in SAR images are often not clear enough, increasing the difficulty of detection, and SAR images may contain various complex terrains and backgrounds, making it difficult to distinguish targets from backgrounds. In addition, targets in SAR images may have different sizes, shapes and directions, increasing the complexity of detection.

[0004] In view of these challenges, researchers have made various explorations. For example, through image preprocessing, including traditional wavelet shrinkage, spatial filtering, frequency domain filtering and other techniques, as well as emerging deep learning denoising methods, these methods aim to reduce noise in SAR images and improve image quality. Researchers have tried various feature extraction methods to better capture the features of targets in SAR images, and have improved detection algorithms (such as improving target positioning algorithms) to improve small target detection capabilities. In recent years, various deep learning models have also been applied to SAR image target detection, trying to improve the accuracy and robustness of detection.

[0005] Despite the continuous development of SAR image target detection technology, existing methods still have many shortcomings. Existing techniques have limited effectiveness in dealing with the speckle noise unique to SAR images, making it difficult to fully utilize the unique information of SAR images. While some methods attempt to improve the detection ability of different types and scales of targets, the generalization still needs to be improved. Moreover, most current methods treat denoising and target detection as independent steps, lacking a unified end-to-end framework, which not only reduces processing efficiency but also can cause information loss. At the same time, existing target positioning methods have insufficient accuracy when processing SAR images, especially in complex backgrounds or small target cases, and the loss function design fails to fully consider the special properties of SAR images. In addition, the training of target detection models often requires a large amount of labeled data, and obtaining a large-scale labeled SAR image dataset is challenging, which to some extent also limits the further improvement of model performance. SUMMARY

[0006] The present disclosure proposes a SAR target detection model combining a denoising network and an open-set target detection technique, solving the problem of insufficient processing of speckle noise in the current SAR target detection field, while overcoming the limitations of traditional target detection for fixed categories, improving target detection accuracy, and expanding the applicability and robustness of the model.

[0007] According to one embodiment of the present disclosure, a synthetic aperture radar image target detection method is proposed, comprising:

[0008] Obtaining a synthetic aperture radar (SAR) image to be detected;

[0009] Using a denoising network to perform denoising processing on the SAR image to obtain a denoised SAR image;

[0010] Establishing an integrated target detection model, the integrated target detection model comprising a denoising network and an open-vocabulary target detection model connected after the denoising network, and training the integrated target detection model using an improved bounding box regression loss function constructed based on corner distance and side overlap information;

[0011] Inputting the denoised SAR image and text prompts into the trained open-vocabulary target detection model for feature extraction and target detection, and outputting a detection result, the detection result comprising the bounding box and category of the recognized target.

[0012] In some embodiments, the denoising network is a pre-trained AGSDNet model.

[0013] In some embodiments, the base model of the open-vocabulary target detection model is a GroundingDINO model.

[0014] In some implementations, the open vocabulary object detection model employs a pre-trained Swin Transformer for image feature extraction and / or a pre-trained BERT for text feature extraction.

[0015] In some implementations, an improved bounding box regression loss function L is constructed as follows: SSCA :

[0016] L SSCA =L SSO +α*L CD ,

[0017] Among them, L SSO For the scaling-side overlap loss, L CD The corner distance loss is represented by α, where α is the weighting coefficient.

[0018] In some implementations, the scaling-side overlap loss L is obtained based on the following formula. SSO :

[0019] L SSO =-0.5*SO,

[0020] Where SO is the overlap ratio between the predicted bounding box and the ground truth bounding box in both width and height.

[0021] In some implementations, the corner distance loss L is obtained based on the following formula. CD :

[0022] L CD =(D corner1 / D enclose )+(D corner2 / D enclose ),

[0023] Among them, D corner1 D is the distance between the top-left corner of the predicted bounding box and the ground truth bounding box. corner2 D is the distance between the bottom right corner of the predicted bounding box and the ground truth bounding box. enclose It is the diagonal distance of the smallest bounding box that contains both the predicted and ground truth boxes.

[0024] In some implementations, the weighting coefficient α has a value range of 0 < α ≤ 1.

[0025] In some implementations, the ensemble object detection model is trained using the AdamW optimizer, where the learning rate and weight decay of the AdamW optimizer both range from 10. -5 Up to 10 -3 between.

[0026] In some implementations, the text prompts include at least one category.

[0027] In some embodiments, the detection result further includes a confidence level between each identified target and the corresponding category.

[0028] According to one embodiment of the present disclosure, a synthetic aperture radar image target detection device is provided, comprising:

[0029] a SAR image acquisition unit configured to acquire a synthetic aperture radar (SAR) image to be detected;

[0030] a denoising unit configured to perform denoising processing on the SAR image using a denoising network to obtain a denoised SAR image;

[0031] an integrated model construction unit configured to establish an integrated target detection model, wherein the integrated target detection model comprises the denoising network and an open-vocabulary target detection model connected after the denoising network, and to train the integrated target detection model using an improved bounding box regression loss function constructed based on corner distance and side overlap information;

[0032] an inference unit configured to input the denoised SAR image and a text prompt into the trained open-vocabulary target detection model to perform feature extraction and target detection, and output a detection result, wherein the detection result includes a bounding box and a category of an identified target.

[0033] According to one embodiment of the present disclosure, an electronic device is provided, comprising a memory and a processor, wherein the memory is configured to store computer instructions executable on the processor, and the processor is configured to implement the method of any one of the above when executing the computer instructions.

[0034] According to one embodiment of the present disclosure, a computer-readable storage medium is provided, which stores a computer program executable by a processor, wherein the program is configured to implement the method of any one of the above when executed by the processor.

[0035] The SAR image target detection method provided by the present disclosure realizes significant performance improvement by combining a denoising network and an open-vocabulary target detection model. First, a special denoising network is introduced to effectively reduce the speckle noise in the SAR image and improve the image quality, laying a good foundation for subsequent target detection. At the same time, the open-vocabulary target detection model greatly expands the range of detectable target categories, enabling the model to flexibly cope with various types of targets and greatly improving the flexibility and applicability of detection. The combination of the denoising network and the open-vocabulary target detection model not only overcomes the limitations of traditional methods on fixed categories, but also enhances the model's ability to handle complex scenes. The present disclosure also introduces an improved bounding box regression loss function, which significantly improves the accuracy of target positioning by considering the corner distance and side overlap information. In particular, it has good performance when dealing with targets with blurred boundaries in SAR images. In addition, the end-to-end integration scheme proposed by the present disclosure not only improves the processing efficiency, but also reduces the information loss that may be caused by intermediate steps.

[0036] Other features and advantages of the present disclosure are described in detail below. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure, together with the description.

[0038] Figure 1 A flowchart of a synthetic aperture radar image target detection method according to one embodiment of the present disclosure is shown.

[0039] Figure 2 The overall architecture of the DGD (Denoise Grounding DINO) model according to one exemplary embodiment of the present disclosure is shown.

[0040] Figure 3 A comparison diagram of an image before denoising and an image after denoising according to one exemplary embodiment of the present disclosure is shown.

[0041] Figure 4 A comparison diagram of ship detection results obtained by the GD model and ship detection results obtained by the DGD model according to one exemplary embodiment of the present disclosure is shown.

[0042] Figure 5 A comparison diagram of car detection results obtained by the GD model and car detection results obtained by the DGD model according to one exemplary embodiment of the present disclosure is shown.

[0043] Figure 6A comparison diagram showing the tank detection result obtained by the GD model and the tank detection result obtained by the DGD model according to one example embodiment of the present disclosure is shown.

[0044] Figure 7 is a structural diagram of an electronic device according to at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0045] The example embodiments will be described in detail herein with reference to the attached drawings. In the following description, unless otherwise indicated, like numbers refer to like elements throughout the description and drawings. The following description is not meant to limit the embodiments of the present disclosure to any particular embodiment described. Rather, the following description is meant to provide an example embodiment of the present disclosure in accordance with the claims set forth below.

[0046] Embodiments of the present disclosure can be applied to a computer system / server, which can operate in a distributed cloud computing environment. A representative example of a distributed cloud computing environment is shown in Fig. 1. In this example, the computer system / server 12 is connected to the network 14 through a network interface or adapter 16. In this example, the computer system / server 12 provides software NMS (Network Management System) 18, which can be used to manage the network 14. The computer system / server 12 can also provide other software applications, such as a web server, a file server, a media server, a database server, etc.

[0047] The computer system / server can be described in the general context of computer system-executable instructions, such as program modules, being executed by a computer system. Generally, program modules can include routines, programs, objects, components, logic, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer system / server can operate in a distributed cloud computing environment where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules can be located in local or remote computer system storage media including storage devices.

[0048] Figure 1 A synthetic aperture radar image target detection method according to an embodiment of the present disclosure is shown. As shown in the figure, the method includes steps 1-4.

[0049] Step 1, obtaining a synthetic aperture radar (SAR) image to be detected.

[0050] A target area can be scanned using a synthetic aperture radar system (SAR system) to collect a SAR image. SAR systems can be mounted on satellites, airplanes, or other platforms. SAR systems can emit microwave signals and receive signals reflected back from the ground to form high-resolution two-dimensional images through complex signal processing techniques.

[0051] Raw SAR images can be pre-processed through geometric correction, radiometric correction, etc., such as to eliminate system errors and atmospheric effects. Processed SAR images can be stored in digital formats, common formats including TIFF, JPEG2000, etc.

[0052] In some real-time monitoring systems, SAR images can be directly transmitted from the radar system to the processing unit, while in research environments, pre-collected SAR image datasets are often used. In some embodiments, suitable SAR images can be selected from an image library according to specific detection tasks. Selection criteria can include image resolution, coverage area, imaging time, etc.

[0053] Step 2, denoising the SAR image using a denoising network to obtain a denoised SAR image.

[0054] Currently, many SAR target detection models only use a single feature extraction and detection network, and little consideration is given to denoising. The present embodiment introduces denoising into target detection to optimize extracted features and improve detection performance.

[0055] In some preferred embodiments, the denoising network is a pre-trained AGSDNet model.

[0056] The inventors have found through in-depth research and comparison that the AGSDNet model combines dilated convolution and gradient information modules as a denoising design, and adding an AGSDNet denoising network before the image encoder has good effects on removing speckle noise in SAR images.

[0057] Specifically, the AGSDNet model is a convolutional neural network composed of a gradient calculation module (GCB), a feature denoising module (FDB), and two attention modules. The feature denoising module includes a convolutional layer and a residual connection block, which includes an average pooling and a convolutional layer to simulate median filter denoising.

[0058] The AGSDNet model can be pre-trained using SAR images, so that the network learns how to effectively remove noise in SAR images. When training the integrated target detection model later, the AGSDNet model initialized by the pre-trained weights is frozen, i.e., the weights of the AGSDNet model are not updated during the training process.

[0059] Step 3, an integrated target detection model is established, which includes a denoising network and an open-vocabulary target detection model connected after the denoising network, and an improved bounding box regression loss function is constructed based on corner distance and side overlap information, and the integrated target detection model is trained.

[0060] The inventors believe that, compared with traditional closed-set target detection models, the open-vocabulary target detection model introduces language information, trains a visual-linguistic alignment space through image-text pairs, has zero-shot learning ability and stronger fine-tuning ability, and can handle a wider range of target categories.

[0061] The target detection model integrating the denoising network and the open-vocabulary target detection model connected after the denoising network includes two main input branches: SAR images and text prompts. The target detection model includes the following key parts:

[0062] Denoising network: located before the image encoder, used to remove speckle noise in the SAR image;

[0063] Two encoders: used for feature extraction of text and denoised image, respectively;

[0064] Feature enhancement module: used for fusion of cross-modal features;

[0065] Cross-modal query selection module: comprehensive determination of cross-modal query.

[0066] Decoder: outputs the recognized bounding box and class results.

[0067] By combining the denoising network with the open-vocabulary target detection model, an end-to-end integrated target detection model is formed. This integration enables the model to handle both denoising of SAR images and open-vocabulary target detection tasks, significantly improving overall detection performance. The model can effectively handle noise problems in SAR images while taking advantage of open-vocabulary target detection to achieve more flexible and accurate target detection.

[0068] In some embodiments, the GroundingDINO model can be selected as the base model of the open-vocabulary target detection model to further improve target detection performance. The feature enhancement module of the GroundingDINO model can fuse image and text features through self-attention and text-image cross-attention mechanisms, the query selection module can use the fused features to comprehensively determine the cross-modal query, and the detection decoder uses self-attention, cross-modal attention, and a feedforward layer to output the final bounding box and class results based on the cross-modal query.

[0069] In some embodiments, a pre-trained Swin Transformer can be employed as an image encoder of the open-vocabulary object detection model for image feature extraction. In some embodiments, a pre-trained BERT can be employed as a text encoder of the open-vocabulary object detection model for text feature extraction.

[0070] In one example, the GroundingDINO model can employ a pre-trained Swin Transformer as an image encoder, employ a pre-trained BERT as a text encoder, and add an AGSDNet model as a denoising network in front of the image encoder to obtain an integrated object detection model Denoise GroundingDINO, referred to as DGD.

[0071] To improve the accuracy of SAR image object detection, the inventors propose an improved bounding box regression loss function, referred to as a scaled SCA (Scaled Side-Corner Aligned, SSCA) loss. The SSCA loss function is constructed based on corner distance and side overlap information, and is used to replace the traditional GIoU loss function. It can better adapt to the characteristics of SAR images and improve the accuracy of target positioning.

[0072] In some embodiments, an improved bounding box regression loss function L SSCA :

[0073] L SSCA = L SSO + α * L CD

[0074] wherein L SSO is a scaled side overlap loss, L CD is a corner distance loss, and α is a weight coefficient. This loss function combines the side overlap information and corner distance information of the bounding box, and can comprehensively evaluate the difference between the predicted box and the ground truth box.

[0075] By combining the scaled side overlap loss L SSO and the corner distance loss L CD , both the overlap degree of the boundary and the accurate positioning of the corner points are considered. Therefore, the bounding box regression loss function L SSC A according to the present embodiment can comprehensively evaluate the quality of the predicted box and improve the overall performance of SAR image object detection.

[0076] The scaled side overlap loss L SSO and the corner distance loss L CD are further discussed below.

[0077] The side overlap ratio SO is defined as follows:

[0078] SO = (w min / w max )+(h min / h max ),

[0079] where w min and h min are the overlapping parts of the prediction box and the ground truth box in width and height, respectively, and w ma x and h max are the union of the two boxes in width and height, respectively. For example, assuming the prediction box is (x1, x2, y1, y2) and the ground truth box is (x'1, x'2, y'1, y'2), we define w min = min(x2, x'2) - max(x1, x'1), w max = max(x2, x'2) - min(x1, x'1), h min = min(y2, y'2) - max(y1, y'1), and h max = max(y2, y'2) - min(y1, y'1).

[0080] The commonly used side overlap loss function is defined as L SO as follows:

[0081] L SO = 2 - SO.

[0082] The inventors have conducted in-depth research, and particularly considered the difference between the range of the classic IoU function (such as GioU) and the range of the conventional L SO , and thus, in some embodiments, a scaled side overlap loss L SSO is obtained based on the following formula as an improved SSO loss:

[0083] L SSO = -0.5 * SO,

[0084] where SO is the overlapping ratio of the prediction box and the ground truth box in width and height.

[0085] The improved L SSO defined in the embodiments has a value range of [-1, 0], which is more easily combined with other loss functions.

[0086] In some embodiments, a corner point distance loss L CD is obtained based on the following formula:

[0087] L CD = (D corner1 / D enclose )+(D corner2 / D enclose ),

[0088] Among them, D corner1 D is the distance between the top-left corner of the predicted bounding box and the ground truth bounding box. corner2 D is the distance between the bottom right corner of the predicted bounding box and the ground truth bounding box. enclose It is the diagonal distance of the smallest bounding box that contains both the predicted and ground truth boxes.

[0089] Through the aforementioned corner distance loss L CD It can effectively measure the alignment between the corner points of the predicted bounding box and the ground truth bounding box.

[0090] Improved bounding box regression loss function L SSCA The weighting coefficient α in L is used to balance L SSO and L CD The contribution of the two loss terms. In some implementations, the weighting coefficient α ranges from 0 < α ≤ 1. In some examples, the value of α may be set to 0.5.

[0091] Step 4: Input the denoised SAR image and text prompt into the trained open vocabulary target detection model for feature extraction and target detection, and output the detection results, which include the bounding boxes and categories of the identified targets.

[0092] The trained open-vocabulary target detection model receives two key inputs: a SAR image processed by a denoising network and a textual cue describing the target, which may be provided by the user. The denoised SAR image has higher quality, with significantly reduced noise. In some implementations, the textual cue may include at least one descriptive word such as "ship" or "aircraft".

[0093] Open-vocabulary object detection models extract features from the received input. For the image portion, the model uses an image encoder (e.g., the Swin Transformer) to extract rich visual features from the denoised SAR image and a text encoder (e.g., the BERT model) to extract semantic features from the text prompts. These two types of features are then fused through a feature enhancement module, which may include self-attention and text-image cross-attention mechanisms, to generate comprehensive cross-modal features.

[0094] After feature fusion, the model's query selection module generates queries using these cross-modal features and feeds them into the decoder. The decoder may include self-attention, cross-attention, and feedforward layers to perform target detection based on these queries. This process fully leverages the combination of visual and semantic information, enabling more accurate target localization and identification in complex SAR images.

[0095] Finally, the model outputs the detection results, including the bounding box of each detected target (representing the target's location and size) and its corresponding category. In some implementations, the model also outputs the confidence score between the identified target and its corresponding category, and different confidence intervals can be obtained by changing the confidence score threshold.

[0096] The SAR image target detection method proposed in this embodiment achieves significant performance improvements by combining a denoising network and an open-vocabulary target detection model. First, a dedicated denoising network effectively reduces speckle noise in SAR images, improving image quality and laying a solid foundation for subsequent target detection. Simultaneously, the use of an open-vocabulary target detection model greatly expands the range of detectable target categories, enabling the model to flexibly handle various target types and significantly improving detection flexibility and applicability. The combination of the denoising network and the open-vocabulary target detection model not only overcomes the limitations of traditional methods on fixed categories but also enhances the model's ability to handle complex scenes. This disclosure also introduces an improved bounding box regression loss function, which significantly improves the accuracy of target localization by comprehensively considering corner distance and side overlap information, especially demonstrating good performance when processing targets with blurred boundaries in SAR images. Furthermore, the end-to-end integration scheme proposed in this disclosure not only improves processing efficiency but also reduces information loss that may occur in intermediate steps.

[0097] In some implementations, the ensemble object detection model is trained using the AdamW optimizer, where the learning rate and weight decay of the AdamW optimizer both range from 10. -5 Up to 10 -3 between.

[0098] In this implementation, the AdamW optimizer was chosen as the core optimization algorithm when training the ensemble object detection model. AdamW is a variant of the Adam (Adaptive Moment Estimation) optimizer, which introduces weight decay regularization on top of the original Adam. Choosing the AdamW optimizer helps improve the efficiency of parameter updates and the model's generalization ability during training.

[0099] Learning rate and weight decay are two key parameters of the AdamW optimizer. The learning rate controls the step size of each parameter update and significantly impacts the model's convergence speed and final performance. Weight decay is a regularization technique that prevents overfitting by adding a parameter norm penalty term to the loss function. In this implementation, both parameters are set to a range of 10. -5 Up to 10 -3 This allows for a balance between ensuring effective model learning and preventing instability or overfitting during the learning process.

[0100] For example, a learning rate within this range can enable the model to learn quickly in the early stage of training and make more detailed parameter adjustments in the later stage. A larger learning rate (e.g., close to 10 -3 ) helps to quickly explore the parameter space in the early stage of training, while a smaller learning rate (e.g., close to 10 -5 ) helps to achieve more fine-grained parameter optimization in the later stage of training. Similarly, a weight decay between 10 -5 and 10 -3 can provide sufficient regularization effect to prevent overfitting, while not overly limiting the learning ability of the model.

[0101] Figure 2 The overall architecture of the DGD (Denoise Grounding DINO) model according to an example embodiment of the present disclosure is shown, which is an integrated model combining denoising and open-vocabulary object detection. The model includes two main input branches: text input and image input.

[0102] The text input branch receives a category hint in the format of "car.ship.aircraft.", which is processed by the text encoder and then passed into the GroundingDINO module. The image input branch is first processed by the denoising network to remove the speckle noise in the SAR image, and then the features are extracted by the image encoder.

[0103] The GroundingDINO module is the core of the model, which includes components such as feature enhancer, query selection module, and decoder. The feature enhancer is used to fuse text and image features, the query selection is used to generate queries based on the fused features, and finally the decoder outputs the detection results. The output of the model includes the bounding box, class, and two losses—SSCA loss and classification loss. The SSCA loss is an improved bounding box regression loss used to optimize the target positioning accuracy. The classification loss is used to evaluate the confidence between the identified target and the corresponding class.

[0104] The following verifies the synthetic aperture radar image object detection scheme according to the present disclosure through specific examples.

[0105] Table 1 dataset description

[0106]

[0107] The SSDD dataset mainly consists of SAR images of small ships, containing a total of 1,160 images. Each image contains an average of 2.1 targets, with an image size of 500*500 pixels, and only the ship class label. The HRSID dataset mainly consists of high-resolution SAR images of ships, containing a total of 5,604 images, with an average of 3.0 targets per image, and only the ship class label. The SARDet-100K dataset collects 10 public SAR datasets and crops each image into a 512*512 image block, including six types of targets: ship, aircraft, car, tank, bridge, and harbor.

[0108] Since the SARDet-100K dataset is very large, the test is very time-consuming, and in this example, a smaller SARDet-100K (partial) dataset is extracted from the original dataset at a ratio of 10:1. First, the class with the most instances in an image can be defined as the class of the image, and then 1,000 images are extracted for each class in the training dataset. Since there are only 837 images of the car class, a total of 5,837 training set images are finally obtained. The test dataset is obtained in the same way.

[0109] In this example, Grounding-DINO-T pre-trained on visible light datasets is used as the initial model, and the SAR pre-trained version of AGSDNet is used to initialize the denoising model, keeping the text encoder and denoising model frozen, and fine-tuning the other structures of the detection model.

[0110] The text input for SSDD and HRSID is "ship.", while the text input for SARDet-100K (partial) is "ship.aircraft.car.tank.bridge.harbor.", and SCA is somewhat robust to α. For simplicity, this example does not consider the impact of different hyperparameters in the loss function. Both SCA loss and SSCA loss take α as 0.5.

[0111] RandomFlip with prob=0.5 and multi-scale RandomChoiceResize are set for training image pre-processing in this example, while fixed 800*1333 resizing is used for test image pre-processing. AdamW is used as the optimizer with a learning rate of 1e-4 and a weight decay of 1e-4; gradient clipping is also used. To keep fairness, all models are trained for 12 rounds in each dataset experiment. All experiments in this example are performed on the MMDetection framework.

[0112] To evaluate the effectiveness of the model, a series of comparative experiments and ablation experiments were conducted on the SSDD, HRSID, and SARDet-100K (partial) datasets to demonstrate the important role of the denoising network, SCA loss function, and improved SSCA loss function. The detailed results are shown in Table 2. In the experiments, AP50 and mAP were compared for each experiment, which are commonly used indicators in target detection.

[0113] Table 2 Performance of different detection models on three SAR target detection datasets

[0114]

[0115] The conclusions are as follows:

[0116] (1) Comparison with SOTA models: The performance of the target detection model GroundingDINO according to the present disclosure is much better than the current SOTA, MSFA (Faster R-CNN+VAN-B), especially on the HRSID dataset, with an increase of about 8.0 in mAP and about 9.0 in AP50, which fully verifies the excellent fine-tuning ability of the open-vocabulary detection model GroundingDINO on the SAR image detection task;

[0117] (2) Comparison of different detection box loss functions: As can be seen from Table 2, there is a significant improvement after using the SCA loss compared to GIoU, and further improvement after using the SSCA loss, and the improvement is more obvious when combined with the denoising network;

[0118] (3) Comparison of whether to add a denoising network: After adding the denoising network, the improvement in mAP is more obvious, indicating that the detection box regression result becomes more accurate.

[0119] In summary, the DGD model according to the present disclosure achieves the best performance on both simple SSDD and HRSD datasets and challenging SAR datasets.

[0120] The detection results of the GroundingDINO (GD) model and the Denoise GroundingDINO (DGD) model can be further compared visually.Figure 3 The left is the image before denoising, and the right is the image after denoising. Figure 4 The left is the ship detection result obtained by the GD model, and the right is the ship detection result obtained according to the DGD model of the present disclosure, the red box represents the prediction result, and the green box represents the true label. Figure 3 And Figure 4 It can be seen that the spots of the ship after denoising are significantly reduced, and the boundary box is more accurate.

[0121] Figure 5 The left in the middle is the car detection result obtained by the GD model, and the right is the car detection result obtained according to the DGD model of the present disclosure. Figure 6 The left in the middle is the tank detection result obtained by the GD model, and the right is the tank detection result obtained according to the DGD model of the present disclosure. Figure 5 And Figure 6 The red box in the middle represents the prediction result, and the green box represents the true label.

[0122] It can be seen that the GD detection picture has two false detections at the bottom and the lower right corner, and from Figure 5 It can be seen that, compared with DGD, GD has three missed detections on the right, which fully embodies the superiority of the DGD model obtained according to the present disclosure. Figure 6 The above examples show that the DGD model performs well in the SAR target detection task. In addition, experimental analysis shows that the integrated target detection model according to the present disclosure is scalable and can be migrated to other detection models, which is conducive to subsequent in-depth research.

[0123] Other details and beneficial effects of the present embodiment are described above, and will not be repeated here.

[0124]

[0125] The electronic device provided by at least one embodiment of the present disclosure includes a memory and a processor, the memory is used to store computer instructions executable on the processor, and the processor is used to implement the synthetic aperture radar image target detection method described in any embodiment or implementation manner of the present disclosure when executing the computer instructions. Figure 7 At least one embodiment of the present disclosure also provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the synthetic aperture radar image target detection method described in any embodiment or implementation manner of the present disclosure.

[0126]

[0127] ​Those skilled in the art will appreciate that one or more embodiments described herein can provide a method, a system or a computer program product. Accordingly, one or more embodiments described herein can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments described herein can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable program code embodied thereon.

[0128] Embodiments in the specification are described in a progressive manner, and the same or similar parts among the embodiments can be mutually referred to. Each embodiment focuses on the difference from other embodiments. In particular, the data processing device embodiment is described simply because it is substantially similar to the method embodiment, and the relevant parts can be referred to the description of the method embodiment.

[0129] The above describes specific embodiments of the specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims can be performed in a different order than the order described in the embodiments and still achieve the desired result. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous or possible.

[0130] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory program carrier for execution by, or to control the operation of, data processing apparatus. Alternatively or additionally, the program instructions can be encoded on an artificially generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus. The computer storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them.

[0131] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit), and / or by programmable data processing apparatuses, which can be portion of hardware processing circuitry that executes specific tasks as described. Apparatuses can also be implemented as a combination of special purpose logic circuitry, e.g., an FPGA or an ASIC, and / or one or more programmed computers.

[0132] Computers suitable for the execution of a computer program include, by way of example, general and / or special purpose microprocessors, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memory and / or a random access memory. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device (e.g., a universal serial bus (USB) flash drive), to name just a few.

[0133] Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.

[0134] While this specification contains many specifics, these should not be construed as limitations on the scope of any invention or on the required scope of patent protection. Certain features outside the scope of the claimed invention are described in this specification, and these should not be construed as excluding those features from the claimed invention. Another aspect concerns implementing various features and / or examples using means-to- perform functions specified in the description. Still another aspect concerns implementing various features and / or examples using other technical and / or functional means.

[0135] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring such order nor limiting of all illustrations to that order, nor requiring that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing can be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0136] Thus, particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. In some cases, actions recited in the claims can be performed in a different order and still achieve desirable results. In addition, the processes depicted in the accompanying figures need not be performed in the particular order described, or in sequential order, to achieve desirable results. In certain implementations, multitasking and parallel processing can be advantageous.

[0137] The above descriptions are only preferred embodiments of one or more embodiments of the present specification, and are not intended to limit one or more embodiments of the present specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of the present specification shall be included in the protection scope of one or more embodiments of the present specification.

Claims

1. A method for target detection in synthetic aperture radar images, characterized in that, include: Acquire synthetic aperture radar (SAR) images of the target object; The SAR image is denoised using a denoising network to obtain a denoised SAR image. The denoising network is a pre-trained AGSDNet model. An integrated object detection model is established, comprising a denoising network and an open-vocabulary object detection model following the denoising network, and an improved bounding box regression loss function constructed based on corner distance and side overlap information is used. Train the integrated object detection model: ， in, For scaling-side overlap loss, For corner distance loss, Here, are the weighting coefficients, , To determine the overlap ratio between the predicted bounding box and the ground truth bounding box in width and height. , This is the distance between the top-left corner of the predicted bounding box and the top-left corner of the ground truth bounding box. This is the distance between the bottom right corner of the predicted bounding box and the ground truth bounding box. The diagonal distance of the smallest bounding box containing both the predicted and ground truth boxes; The denoised SAR image and text prompts are input into the trained open vocabulary target detection model for feature extraction and target detection, and the detection results are output, including the bounding boxes and categories of the identified targets.

2. The synthetic aperture radar image target detection method according to claim 1, characterized in that: The underlying model for the open vocabulary target detection model is the GroundingDINO model.

3. The synthetic aperture radar image target detection method according to claim 1 or 2, characterized in that: The open vocabulary object detection model uses a pre-trained Swin Transformer for image feature extraction and / or a pre-trained BERT for text feature extraction.

4. The synthetic aperture radar image target detection method according to claim 1, characterized in that: The weighting coefficient The range of values ​​is .

5. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The ensemble object detection model was trained using the AdamW optimizer, where the learning rate and weight decay values ​​both ranged from [value missing]. arrive between.

6. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The text prompts include at least one category.

7. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The detection results also include the confidence level between each identified target and its corresponding category.

8. A synthetic aperture radar image target detection device, characterized in that, include: SAR image acquisition unit, used to acquire synthetic aperture radar (SAR) images of the target object; The denoising unit is used to denoise the SAR image using a denoising network to obtain a denoised SAR image. The denoising network is a pre-trained AGSDNet model. An integrated model building unit is used to build an integrated object detection model, which includes a denoising network and an open-vocabulary object detection model following the denoising network, and uses an improved bounding box regression loss function based on corner distance and side overlap information as follows. Train the integrated object detection model: ， in, For scaling-side overlap loss, For corner distance loss, Here, are the weighting coefficients, , To determine the overlap ratio between the predicted bounding box and the ground truth bounding box in width and height. , This is the distance between the top-left corner of the predicted bounding box and the top-left corner of the ground truth bounding box. This is the distance between the bottom right corner of the predicted bounding box and the ground truth bounding box. The diagonal distance of the smallest bounding box containing both the predicted and ground truth boxes; The inference unit is used to input the denoised SAR image and text prompts into the trained open vocabulary target detection model for feature extraction and target detection, and output the detection results, which include the bounding boxes and categories of the identified targets.

9. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store computer instructions that can be executed on the processor, and the processor being used to implement the method of any one of claims 1 to 7 when executing the computer instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method according to any one of claims 1 to 7.

Images