Small sample SAR (Synthetic Aperture Radar) image aircraft detection method based on double-level enhancement adaptation

Abstract

The invention discloses a small sample SAR (Synthetic Aperture Radar) image aircraft detection method based on double-level enhancement adaptation, which performs general knowledge enhancement and domain difference adaptation from an image level and a feature level, and effectively improves the SAR aircraft detection performance under a small sample condition. In order to realize effective knowledge migration, three stages of image conversion, pre-training and fine tuning are designed in the migration learning training process of the target detection network. In an image conversion stage, CycleGAN is used to convert an optical image into a generated image in an SAR style, and edge enhancement and histogram loss are introduced to eliminate image level differences. In a pre-training stage, a target detection model is used to complete aircraft target detection training on a generated image, and representative feature enhancement and decoupling field matching are introduced to eliminate feature level differences. And finally, in the fine tuning stage, SAR field learning is completed on a real SAR image based on the target detection model in the pre-training stage, and high-precision SAR aircraft detection is realized.

Description

Technical Field

[0001] This invention belongs to the field of SAR image processing technology, specifically relating to a method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation. Background Technology

[0002] Aircraft target detection can quickly acquire the location and quantity distribution of various aircraft targets in airport areas, playing a vital role in urban logistics management, emergency disaster relief, and other fields. Thanks to the unique active imaging mechanism of Synthetic Aperture Radar (SAR), SAR image aircraft detection provides a stable technical means for acquiring aircraft information unaffected by weather and lighting conditions. In recent years, deep learning-based target detection methods have demonstrated excellent detection performance in SAR image aircraft detection. Their automatic target feature extraction learning mode effectively solves the problems of complex SAR aircraft scattering characteristics and severe airport background interference. Since mainstream target detection networks rely on a fully supervised learning paradigm, their training process often requires a large number of labeled samples. However, due to the special nature of aircraft targets, obtaining a large number of labeled SAR aircraft samples is often difficult. On the one hand, stationary aircraft targets are basically only found within the airport area, and strict regional restrictions result in a limited number of SAR images containing targets. On the other hand, aircraft targets in SAR images often appear as discrete, strongly scattering points, whose shapes do not conform to human visual habits, making manual annotation complex and time-consuming. Therefore, how to achieve high-accuracy aircraft detection in SAR images under conditions of scarce sample size has become a key scientific issue driving the development of SAR remote sensing interpretation technology.

[0003] To address the aircraft detection problem in SAR images with scarce samples, existing techniques propose transferring general target detection knowledge from optical images, which are more abundant and have lower annotation costs, to improve detection performance. Among these, adversarial learning methods based on global feature matching are the mainstream approach. Specifically, global feature matching aligns the feature distributions of the backbone network between optical and SAR images during the training process, thereby promoting the extraction of general target detection knowledge shared between the two domains and improving transfer learning effectiveness. Its learning method typically relies on an additional domain classifier added to the feature map, which includes a gradient reversal layer (GRL) and two layers... The network consists of a convolutional layer and a normalization function layer. The gradient reversal layer does not modify the input feature map during forward inference, but multiplies the gradient value by a normalization function during the backpropagation phase of network training. This approach enables end-to-end adversarial learning between the domain classifier and the detection network. Two convolutional layers are used to classify the features of the backbone network, and the domain classification results are output through a normalization function. Through domain adversarial learning, this method can effectively reduce the global domain difference between optical and SAR images at the feature level, and improve the transfer learning effect of optical image knowledge to SAR target detection tasks.

[0004] While existing domain adversarial learning methods can effectively reduce global domain differences, global feature matching struggles to achieve effective knowledge transfer due to the significant image style differences between optical and SAR images and the complex scattering structure of SAR aircraft. Specifically, the differences between optical and SAR images manifest in variations in overall brightness, contrast, texture, and target structure. Optical images contain multiple color channels, generally exhibiting weak contrast, and clearly show the structural outlines and detailed textures of aircraft targets. In contrast, SAR images often contain only a single scattering intensity channel, with significant differences in scattering intensity between foreground targets and background regions. Aircraft targets appear as discretely distributed strong scattering points, lacking clear outlines and texture features. When using global distribution adaptation to adjust feature distribution, it cannot effectively address the interference caused by the overall image style differences between optical and SAR images on the detector's shallow feature extraction. Furthermore, due to the significant differences in aircraft target morphology, global feature matching struggles to accurately match foreground target features between optical and SAR images, potentially leading to mismatches between foreground target features and background region features, resulting in erroneous knowledge transfer. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for aircraft detection in small-sample SAR images based on two-level enhancement adaptation. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for aircraft detection in small-sample SAR images based on two-level enhancement adaptation, wherein the small-sample SAR image aircraft detection includes: Acquire the SAR image to be detected; The SAR image to be detected is input into the trained target detection model to obtain the detection result; among which, The transfer learning training process of the object detection model includes: image conversion stage, pre-training stage, and fine-tuning stage; The image conversion stage includes: constructing a source domain dataset and a target domain dataset, where the source domain dataset consists of optical images and the target domain dataset consists of SAR images; performing edge enhancement on the optical images in the source domain dataset; training the initial CycleGAN model based on the target domain dataset and the edge-enhanced source domain dataset, using adversarial loss, histogram loss, and cycle consistency loss to obtain a trained CycleGAN model; and inputting the edge-enhanced source domain dataset into the trained CycleGAN model to generate a generator that converts optical images to SAR images, thus obtaining the generated domain dataset. The pre-training phase includes: pre-training the initial object detection model based on the generator domain dataset and the target domain dataset, and based on object detection loss, representative feature enhancement loss and decoupling domain matching loss to obtain a pre-trained object detection model; The fine-tuning phase includes: training the pre-trained object detection model based on the target domain dataset and the object detection loss to obtain the trained object detection model.

[0006] The beneficial effects of this invention are: This invention proposes a small-sample SAR image aircraft detection method based on dual-level enhancement and adaptation. It enhances general knowledge and adapts to domain differences at both the image and feature levels, effectively improving SAR aircraft detection performance under small-sample conditions. To achieve effective knowledge transfer, this invention designs three stages in the transfer learning training process of the target detection network: image transformation, pre-training, and fine-tuning. In the image transformation stage, CycleGAN is used to convert optical images into SAR-style generated images, and edge enhancement and histogram loss are introduced to eliminate image-level differences. In the pre-training stage, the target detection model is trained on the generated images to perform aircraft target detection, and representative feature enhancement and decoupled domain matching are introduced to eliminate feature-level differences. Finally, in the fine-tuning stage, the target detection model from the pre-training stage is used to complete SAR domain learning on real SAR images, achieving high-precision SAR aircraft detection. By using the method proposed in this invention, small-sample SAR aircraft detection can effectively combine optical image knowledge and achieve higher detection accuracy.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a small-sample SAR image aircraft detection method based on dual-level enhancement adaptation provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the transfer learning training framework model provided in an embodiment of the present invention; Figure 3This is a schematic diagram of the histogram loss calculation process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the publicly available dataset provided in an embodiment of the present invention; Figure 5 These are schematic diagrams illustrating the visualization results of SAR image aircraft detection using different methods provided in the embodiments of the present invention. Figure 6 This is a schematic diagram of the ablation experiment image generation results at different stages provided in the embodiments of the present invention. Detailed Implementation

[0009] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0010] Please see Figure 1 This invention provides a method for aircraft detection in small-sample SAR images based on two-level enhancement adaptation, specifically including the following steps: S10. Acquire the SAR image to be detected.

[0011] In this embodiment of the invention, a SAR image containing aircraft targets in the airport area is obtained as the SAR image to be detected. Alternatively, a SAR image can be randomly obtained from the publicly available SAR-aircraft-1.0 dataset as the SAR image to be detected.

[0012] S20. Input the SAR image to be detected into the trained target detection model to obtain the detection result.

[0013] To improve SAR aircraft detection accuracy under small sample conditions, this invention proposes a transfer learning training framework model for target detection, such as... Figure 2 As shown, the transfer learning training process of this object detection model includes: image conversion stage, pre-training stage, and fine-tuning stage; among which, The image conversion stage includes: constructing a source domain dataset and a target domain dataset, where the source domain dataset consists of optical images and the target domain dataset consists of SAR images; performing edge enhancement on the optical images in the source domain dataset; training an initial CycleGAN model based on the target domain dataset and the edge-enhanced source domain dataset, using adversarial loss, histogram loss, and cycle consistency loss to obtain a trained CycleGAN model; and inputting the edge-enhanced source domain dataset into the trained CycleGAN model to generate a generator that converts optical images to SAR images, resulting in the generated domain dataset. More specifically: This invention collects optical images and SAR images containing aircraft targets in different airport areas to construct a training dataset, wherein the optical images serve as the source domain dataset. SAR images as target domain datasets .

[0014] To enhance the spatial structure information and suppress the color texture information in optical images, embodiments of the present invention use a Sobel filter on the source domain dataset. Edge enhancement is performed on optical images. It employs two differential filters, one horizontal and one vertical, to convolve the optical image, and then synthesizes the gradients in both directions to obtain the final edge information. For optical images in the source domain dataset... Edge enhancement is performed to obtain an edge-enhanced optical image. The calculation process is as follows: ; in, Represents the convolution operation. These are differential filters for the horizontal and vertical directions, respectively, in the following forms: ; By analyzing the source domain dataset Edge enhancement is performed on all optical images to obtain the source domain dataset for edge enhancement. .

[0015] This invention uses the classic image style transfer method CycleGAN as the baseline model. The CycleGAN model includes two identical generator models, namely the first generator. Second generator These are used to implement style transfer from optical images to SAR images and from SAR images to optical images, respectively, both based on the classic U-net network architecture; two identical discriminator models, namely the first discriminator. Second discriminator These are used to identify whether the input image is a SAR image or an optical image, respectively, both based on the classic PatchGAN architecture.

[0016] In the target domain dataset Edge-enhanced source domain dataset Before training the initial CycleGAN model, the following steps are taken: First, train the target domain dataset. Source domain datasets of SAR images and edge enhancement Normalization processing is performed on optical images. For example, from edge-enhanced source domain datasets. Randomly selected optical images SAR images are randomly selected from the target domain dataset. ,Will and Perform a normalization operation to normalize its pixel value range to Within the scope. Target domain dataset Edge-enhanced source domain dataset All images in the dataset are normalized and then used as input data.

[0017] Based on the normalized target domain dataset and the edge-enhanced source domain dataset, and trained using adversarial loss, histogram loss, and cycle consistency loss, a trained CycleGAN model is obtained. This model includes: a first generator converts optical images in the edge-enhanced source domain dataset to SAR images, and converts the output of the second generator to SAR images; a second generator converts SAR images in the target dataset to optical images, and converts the output of the first generator to optical images; a first discriminator, based on the target domain dataset, determines whether the output of the first generator is a SAR image; a second discriminator, based on the edge-enhanced source domain dataset, determines whether the output of the second generator is an optical image; and so on. Based on the discrimination results of the first and second discriminators, the adversarial loss is calculated; based on the SAR image in the target domain dataset and the edge-enhanced optical image in the source domain dataset, and the transformation results of the first and second generators on the edge-enhanced optical image in the source domain dataset and the SAR image in the target dataset, respectively, the histogram loss is calculated; based on the SAR image in the target dataset and the edge-enhanced optical image in the source domain dataset, and the transformation results of the first and second generators on the output of the second generator and the output of the first generator, respectively, the cycle consistency loss is calculated; based on the adversarial loss, histogram loss, and cycle consistency loss, the initial CycleGAN model is trained to obtain the trained CycleGAN model. The formula for calculating the adversarial loss is as follows: ; ; in, This represents the adversarial loss of the generator. This refers to the SAR image obtained by the first generator from the optical image of the source domain dataset with edge enhancement, which can be called a spurious SAR image. This refers to the optical image generated by the second generator from the SAR images in the target dataset, which can be called a fake optical image. This indicates that the first discriminator determines whether the output of the first generator is the discrimination result of the SAR image. This indicates that the second discriminator determines whether the output of the second generator is an optical image. This indicates the adversarial loss of the discriminator. Represents the optical image in the source domain dataset for edge enhancement. This indicates the discrimination result of the second discriminator in determining whether an optical image in the source domain dataset with edge enhancement is an optical image. Represents SAR images in the target dataset. This indicates the identification result of the first discriminator in determining whether a SAR image in the target dataset is a SAR image.

[0018] like Figure 3 As shown, histogram loss constructs differentiable histogram features using a kernel density estimation method, and quantitatively measures the distribution difference between generated and real images by calculating the histogram distance between unpaired images. For images containing pixels... Image Using kernel functions The probability density function is estimated as follows: ; in, These are parameters that control the smoothness of the kernel function. Kernel function Choose to use the Sigmoid function. The derivative of is calculated as follows: ; Image pixel value distribution range Divided into There are sub-intervals, each of which is Based on this sub-interval division, the image The histogram features are represented as follows: ; Each element Representing the The pixel distribution intensity across each sub-interval is calculated as follows: ; in, These represent the left and right boundaries of the sub-intervals, respectively. Histogram loss. The squared earth mover's distance (SEMD) is used to quantify the characteristics of false optical images. With real optical images False SAR images Compared with real SAR images The histogram distance between them is calculated as follows: ; in, express Histogram features, express Histogram features, express Histogram features, express Histogram features, It represents the squared distance of a seismic event. Indicates the first weighted parameter; squared seismic distance. The calculation method is as follows: ; Cyclic consistency loss The calculation is expressed by the formula: ; in, This represents the optical image obtained by the second generator from the output of the first generator. This represents the SAR image obtained by the first generator converting the output of the second generator. This represents the second weighting parameter. This indicates the calculation of the L1 norm.

[0019] Based on adversarial losses Histogram loss and cycle consistency loss Calculate the first generator in the CycleGAN model Second generator First Discriminator Second discriminator The parameter gradients are used to update the first generator using the Adam optimizer. Second generator First Discriminator Second discriminator Model parameters.

[0020] The process of training the CycleGAN model with updated parameters, based on the target domain dataset and the source domain dataset with edge enhancement, is repeated until the training data has been traversed. Next, complete the model training and save the first generator. Parameter weights .

[0021] Load the saved first generator Parameter weights Edge-enhanced source domain dataset All optical images are input into the first generator. Obtain the corresponding SAR style pseudo-image To obtain the generated domain dataset .

[0022] Furthermore, the pre-training phase includes: pre-training the initial object detection model based on the generator domain dataset and the target domain dataset, and using object detection loss, representative feature enhancement loss, and decoupling domain matching loss to obtain a pre-trained object detection model. More specifically: This invention uses the classic object detection model YOLO, such as YOLOv5, as the benchmark detection model. It consists of three parts: a backbone network, a feature pyramid structure, and a detection head. Before pre-training the initial target detection model, the following steps are performed: normalizing the SAR images in both the generator domain dataset and the target domain dataset. For example, from the generator domain dataset... Randomly selected optical images Get its corresponding tag From the target domain dataset Randomly selected SAR images Get its corresponding tag Among them, the tags Label the aircraft target in the corresponding image, including the bounding box coordinates and its length and width parameters. and Perform a normalization operation to normalize its pixel value range to Within the scope. Target domain dataset Generate domain dataset All images in the dataset are normalized and then used as input data.

[0023] Based on the normalized generator domain dataset and target domain dataset, and using target detection loss, representative feature enhancement loss, and decoupling domain matching loss, an initial target detection model is pre-trained to obtain a pre-trained target detection model. This includes: inputting optical images from the generator domain dataset into the initial target detection model; calculating the target detection loss based on the detection results and corresponding labels output by the initial target detection model; calculating the corresponding foreground target mask based on the labels corresponding to the optical images in the generator domain dataset and the labels corresponding to the SAR images in the target domain dataset; and inputting the optical images from the generator domain dataset and the SAR images from the target domain dataset into the backbone network of the target detection model to obtain the corresponding optical features. Optical and SAR features are input into a feature discriminator to obtain the corresponding optical foreground target prediction probabilities and SAR foreground target prediction probabilities. Based on the foreground target mask, optical foreground target prediction probabilities, and SAR foreground target prediction probabilities, the corresponding representative feature enhancement loss is calculated. Optical and SAR features are simultaneously input into two different neighborhood classifiers for the foreground and background, respectively, to obtain the corresponding neighborhood classification probabilities. Based on the foreground target mask and neighborhood classification probabilities, the corresponding decoupling neighborhood matching loss is calculated. The initial target detection model is pre-trained based on the target detection loss, representative feature enhancement loss, and decoupling neighborhood matching loss to obtain a pre-trained target detection model. The target detection loss in this embodiment of the invention is calculated by generating a domain dataset. Medium optical images Input the target detection model, based on the detection results and corresponding aircraft target tags Calculate the target detection loss, labeled as The calculation of the specific target detection loss is described in existing technology and will not be repeated here.

[0024] Generate domain dataset Medium optical images and target domain dataset SAR images Input the backbone network of the object detection model respectively The corresponding optical characteristics are obtained. and SAR features .

[0025] For optical features and SAR features Based on the corresponding optical and SAR aircraft target tags and Calculate the corresponding foreground target mask and The calculation is based on determining whether the geometric position of each pixel in the feature is located within the bounding box region corresponding to the aircraft target label. The specific calculation method is as follows: ; ; in, These represent the horizontal and vertical coordinates of each pixel in the feature. Using this foreground object mask, pixels in the foreground and background regions can be distinguished, enabling refined feature enhancement and matching. Specifically, the foreground object mask... and The number of optical foreground target features, optical background region features, SAR foreground target features, and SAR background region features can be calculated and denoted as follows: .

[0026] To enhance the ability of the object detection model to distinguish between foreground object features and background region features during the pre-training stage, and to extract representative foreground object features, this embodiment of the invention adds a feature discriminator to the output of the backbone network in the object detection model. The feature discriminator distinguishes foreground objects from background regions based on the output of the backbone network. This feature discriminator consists of two layers. The convolutional layers output the predicted probability of the foreground object based on the corresponding pixel features. The calculation of the representative feature enhancement loss is expressed by the formula: ; ; in, This represents the loss in representative feature enhancement of an optical image. These represent the pixel width and height of the input feature, respectively. Represents the number of pixels in the optical image in the generated domain dataset. Foreground target mask, Indicates a feature discriminator, Represents the number of pixels in the optical image in the generated domain dataset. The optical features obtained from the backbone network of the target detection model express The predicted probability of the foreground target output after passing through the feature discriminator. This represents the representative feature enhancement loss of the SAR image. Represents the pixels of the SAR image in the target domain dataset. Foreground target mask, Represents the pixels of the SAR image in the target domain dataset. SAR features obtained from the backbone network in the target detection model express The predicted probability of the foreground target output after passing through the feature discriminator.

[0027] To reduce the difference between foreground targets and background regions in the optical-SAR domain during the pre-training stage and improve the transferability of feature learning, this embodiment of the invention adds two parallel neighborhood classifiers to the output of the backbone network in the target detection type. and The distribution of foreground target features and background region features output by the backbone network is adapted between optical and SAR methods. The domain classifier consists of a gradient inversion layer and two layers. The system consists of convolutional layers, and the output is the neighborhood classification probability for each corresponding pixel. This is followed by a decoupled neighborhood matching loss derived from a neighborhood classifier for the foreground. The calculation is expressed by the formula: ; in, Indicates the number of features of the optical foreground target. Indicates the number of SAR foreground target features. This represents a domain classifier for the foreground. express The domain classification probability output by the domain classifier applied to the foreground. express The domain classification probability output after passing through a domain classifier for the foreground; Decoupling the domain matching loss from the background domain classifier The calculation is expressed by the formula: ; in, Indicates the number of optical background target features. Indicates the number of SAR background target features. This represents a domain classifier for the context. express The domain classification probability output by the domain classifier applied to the background. express The domain classification probability output after passing through a domain classifier for the background.

[0028] Based on target detection loss Representativeness enhancement loss Matching loss in decoupling domain Calculate the target detection model The parameter gradients are used to update the model parameters of the object detection model using the SGD optimizer.

[0029] The process of training an object detection model with updated parameters, based on the generated domain dataset and the target domain dataset, and using object detection loss, representative feature enhancement loss, and decoupling domain matching loss, is repeated until the training data is traversed. Next, complete the training of the object detection model and save the object detection model. The parameter weights are denoted as .

[0030] Furthermore, the fine-tuning stage includes: training the pre-trained object detection model based on the target domain dataset and using the object detection loss to obtain a trained object detection model. More specifically: In this embodiment of the invention, the fine-tuning stage involves fine-tuning the pre-trained target detection model. Specifically, the baseline detection model is D, and the saved parameter weights of the target detection model are loaded. This is used as the initialization parameter weight of the target detection model at this time.

[0031] Before training the pre-trained target detection model, the following steps are performed: normalizing the SAR images in the target domain dataset. For example, from the target domain dataset... Randomly selected SAR images Get its corresponding tag ,right Perform a normalization operation to normalize its pixel value range to Internal. Target domain dataset All images in the dataset are normalized and used as input data.

[0032] A trained object detection model is obtained by training a pre-trained object detection model using a normalized target domain dataset and an object detection loss. The object detection loss is calculated by using the target domain dataset... SAR images Input the target detection model, based on the detection results and corresponding aircraft target tags Calculate the target detection loss, labeled as The calculation of the specific target detection loss is described in existing technology and will not be repeated here.

[0033] Based on target detection loss Calculate the parameter gradients of the object detection model D, and update the model parameters of the object detection model using the SGD optimizer.

[0034] Repeatedly train the object detection model with updated parameters based on the target domain dataset and the object detection loss until all training data has been traversed. Next, complete the target detection model. Training involves saving the parameter weights of the object detection model, denoted as... This refers to the trained object detection model.

[0035] As can be seen, this invention effectively overcomes the inherent differences between optical and SAR by designing image-level and feature-level knowledge transfer methods, thereby improving the accuracy of SAR aircraft detection. Compared with existing technologies, this invention eliminates optical-SAR image differences at two levels: image-level and feature-level, achieving a more refined knowledge transfer effect. For image-level differences, this invention uses the CycleGAN style transfer network as a benchmark and introduces edge enhancement and histogram loss to improve the similarity between the generated image and the real SAR image. For feature-level differences, this invention introduces representative feature enhancement and decoupled neighborhood matching in the target detection network to enhance the feature differences between foreground targets and background regions and eliminate neighborhood differences between optical and SAR, thereby improving the generalization ability of the model representation.

[0036] Furthermore, after the aforementioned image conversion stage, pre-training stage, and fine-tuning stage, the trained target detection model is determined. The SAR image to be detected is then normalized, and the input is normalized to... Inside, load and save the parameter weights. The corresponding target detection model takes the normalized SAR image to be detected as input and outputs the detection result.

[0037] To verify the effectiveness of the small-sample SAR image aircraft detection method based on dual-level enhancement adaptation provided in this embodiment of the invention, the following experiments were conducted.

[0038] 1. Experimental conditions The experimental data used partial optical images from the publicly available HRRSD dataset and SAR images from the SAR-aircraft-1.0 dataset as the transfer learning dataset, such as... Figure 4 As shown, the left side (a) illustrates the optical image, and the right side (b) illustrates the SAR image. The optical images were selected from the corresponding original dataset, with images containing aircraft targets, resulting in 1970 images as the source domain dataset; the SAR images were selected from the corresponding original dataset, with 552 / 427 images selected as the target domain training / test dataset.

[0039] Regarding the experimental setup, the number of training rounds in the image conversion stage... The optimizer learning rate is set to 60, with a learning rate of 0.0002, which decays linearly to 0 starting from the 30th epoch; this represents the number of training epochs during the pre-training and fine-tuning phases. and The value is 300, the optimizer learning rate is 0.01, and it decays to 0.0001 according to a cosine curve. The kernel width of the histogram loss is... Number of subintervals Loss weight Set them to 3.0 and 10.0 respectively.

[0040] 2. Experimental Results 2.1 Comparative Experiment Based on the above dataset, experiments on optical-SAR transfer learning for aircraft detection were conducted using different transfer learning methods. The experimental results are shown in Table 1. It can be seen that compared to the baseline model without transfer learning methods, the detection accuracy of each model is significantly improved after using transfer learning methods. Furthermore, the overall performance of the YOLO-based detector is superior to that of the Faster R-CNN-based detector. Compared to other transfer learning methods, the method proposed in this invention achieves the highest SAR image aircraft detection accuracy, demonstrating its ability to achieve efficient optical-SAR knowledge transfer.

[0041] Table 1 Comparison of results from different transfer learning methods

[0042] In Table 1, the comparative experiments based on the Faster R-CNN detector show the following: the first row represents the detection methods in Faster R-CNN that did not employ transfer learning; the second to fourth rows represent the detection methods based on DAF (Domain Adaptive Faster R-CNN), SWDA (Strong-Weak Distribution Alignment), and HTCN (Hierarchical Transferability Calibration Network), respectively. Similarly, the comparative experiments based on the YOLO detector show the first row represents the detection methods in YOLO that did not employ transfer learning; the second to fourth rows represent the detection methods based on MS-YOLO (Multiscale Domain Adaptive YOLO), SF-YOLO (Source-Free YOLO), and DHEA (the method proposed in this invention), respectively. The comparison metrics P, R, and F1 represent precision, recall, and F1-score, respectively; AP is the area under the PR curve, with a larger area indicating better model performance. 50 The AP value is calculated when the threshold for Intersection over Union (IoU) is fixed at 0.5.

[0043] Figure 5 The visualization results of aircraft detection in SAR images using different methods are presented. The target detection model used in the detection is the YOLOv5 detector. Figure 5In the diagram: (a) shows the real labels for comparison; (b) shows the detection results using only the YOLOv5 detector without transfer learning; (c) shows the detection results using MS-YOLO with transfer learning; (d) shows the detection results using SF-YOLO with transfer learning; and (e) shows the detection results using the method of this invention. The blue, green, red, and yellow rectangles represent the real target, correct detection, false alarm detection, and missed detection results, respectively. It can be seen that the method proposed in this invention has superior performance in identifying background interference and accurately locating aircraft targets.

[0044] 2.2 Ablation Experiment To verify the effectiveness of each stage proposed in this invention, stage ablation experiments were conducted, and the results are shown in Table 2. It can be seen that compared to directly using optical image pre-training, introducing edge enhancement and histogram loss can effectively improve the accuracy of aircraft detection after transfer learning. By further introducing representative feature enhancement and decoupled neighborhood matching in the pre-training stage, the model accuracy can be further improved, ultimately achieving an AP of 0.629. 50 With an AP of 0.337, these represent improvements of 5.4% and 3.8% respectively compared to the baseline model.

[0045] Table 2. Validation experiments at each stage

[0046] also, Figure 6 The image generation results are shown when using different stages. Figure 6 In the diagram: (a) illustrates the original optical image; (b) illustrates the image generation result after edge enhancement; (c) illustrates the image generation result after CycleGAN; (d) illustrates the image generation result after edge enhancement and CycleGAN; (e) illustrates the image generation result using the complete stage of this invention. Compared to the original optical image, the edge-enhanced image retains the spatial structure information of the original image but suppresses redundant color and texture information. When directly using CycleGAN for style transfer, some errors occur in the generated result, such as the runway area exhibiting bright textures inconsistent with the SAR style. After combining edge enhancement and histogram loss, the generated result shows the aircraft target as a discrete scattering point, exhibiting the highest visual similarity to the real SAR image, proving the effectiveness of the proposed method.

[0047] In summary, the small-sample SAR image aircraft detection method based on dual-level enhancement and adaptation proposed in this invention effectively improves SAR aircraft detection performance under small-sample conditions by performing general knowledge enhancement and domain difference adaptation at both the image level and feature level. To achieve effective knowledge transfer, this invention designs three stages in the transfer learning training process of the target detection network: image transformation, pre-training, and fine-tuning. In the image transformation stage, CycleGAN is used to convert optical images into SAR-style generated images, and edge enhancement and histogram loss are introduced to eliminate image-level differences. In the pre-training stage, the target detection model is used to train aircraft target detection on the generated images, and representative feature enhancement and decoupled domain matching are introduced to eliminate feature-level differences. Finally, in the fine-tuning stage, the target detection model from the pre-training stage is used to complete SAR domain learning on real SAR images, achieving high-precision SAR aircraft detection. By using the method proposed in this invention, small-sample SAR aircraft detection can effectively combine optical image knowledge and achieve higher detection accuracy.

[0048] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0049] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the specification and accompanying drawings, will understand and implement other variations of the disclosed embodiments in carrying out the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0050] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for aircraft detection in small-sample SAR images based on two-level enhancement adaptation, characterized in that, The small-sample SAR image aircraft detection method includes: Acquire the SAR image to be detected; The SAR image to be detected is input into the trained target detection model to obtain the detection result; among which, The transfer learning training process of the object detection model includes: image conversion stage, pre-training stage, and fine-tuning stage; The image conversion stage includes: constructing a source domain dataset and a target domain dataset, where the source domain dataset consists of optical images and the target domain dataset consists of SAR images; performing edge enhancement on the optical images in the source domain dataset; training the initial CycleGAN model based on the target domain dataset and the edge-enhanced source domain dataset, using adversarial loss, histogram loss, and cycle consistency loss to obtain a trained CycleGAN model; and inputting the edge-enhanced source domain dataset into the trained CycleGAN model to generate a generator that converts optical images to SAR images, thus obtaining the generated domain dataset. The pre-training phase includes: pre-training the initial object detection model based on the generator domain dataset and the target domain dataset, and based on object detection loss, representative feature enhancement loss and decoupling domain matching loss to obtain a pre-trained object detection model; The fine-tuning phase includes: training the pre-trained object detection model based on the target domain dataset and the object detection loss to obtain the trained object detection model.

2. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, Before inputting the SAR image to be detected into the trained target detection model, the following steps are included: The SAR image to be detected is normalized.

3. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, Edge enhancement is performed on optical images in the source domain dataset, including: The Sobel filter is used to enhance the edges of the optical images in the source domain dataset.

4. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, Before training the initial CycleGAN model during the image conversion stage, the following steps are included: Normalization was performed on the SAR images in the target domain dataset and the optical images in the edge-enhanced source domain dataset, respectively.

5. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, In the image transformation stage, the CycleGAN model includes a first generator, a second generator, a first discriminator, and a second discriminator. This corresponds to a source domain dataset based on the target domain dataset and edge enhancement data. The trained CycleGAN model is obtained by training the initial CycleGAN model using adversarial loss, histogram loss, and cycle consistency loss. The trained CycleGAN model includes: The first generator converts the optical image in the source domain dataset with edge enhancement to a SAR image, and the output of the second generator is converted to a SAR image; The second generator converts the SAR images in the target dataset into optical images, and also converts the output of the first generator into optical images; The first discriminator, based on the target domain dataset, determines whether the output of the first generator is a SAR image. The second discriminator, based on the edge-enhanced source domain dataset, determines whether the output of the second generator is an optical image; Based on the discrimination results of the first and second discriminators, the adversarial loss is calculated. Based on the SAR image in the target domain dataset and the optical image in the source domain dataset with edge enhancement, as well as the conversion results of the first generator and the second generator on the optical image in the source domain dataset with edge enhancement and the SAR image in the target domain dataset, respectively, the histogram loss is calculated. Based on the SAR image in the target dataset and the optical image in the source domain dataset with edge enhancement, as well as the transformation results of the first generator and the second generator on the output of the second generator and the output of the first generator, respectively, the cycle consistency loss is calculated. The initial CycleGAN model is trained using adversarial loss, histogram loss, and cycle consistency loss to obtain a trained CycleGAN model.

6. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 5, characterized in that, The formula for calculating the adversarial loss is as follows: ； ； in, This represents the adversarial loss of the generator. This represents the SAR image obtained by the first generator from the optical images in the source domain dataset with edge enhancement. This represents the optical image generated by the second generator from the SAR images in the target dataset. This indicates that the first discriminator determines whether the output of the first generator is the discrimination result of the SAR image. This indicates that the second discriminator determines whether the output of the second generator is an optical image. This indicates the adversarial loss of the discriminator. Represents the optical image in the source domain dataset for edge enhancement. This indicates the discrimination result of the second discriminator in determining whether an optical image in the source domain dataset with edge enhancement is an optical image. Represents SAR images in the target dataset. This indicates the identification result of the first discriminator in determining whether a SAR image in the target dataset is a SAR image; Histogram loss The calculation is expressed by the formula: ； in, express Histogram features, express Histogram features, express Histogram features, express Histogram features, It represents the squared distance of a seismic event. Indicates the first weight parameter; Cyclic consistency loss The calculation is expressed by the formula: ； in, This represents the optical image obtained by the second generator from the output of the first generator. This represents the SAR image obtained by the first generator converting the output of the second generator. This represents the second weighting parameter. This indicates the calculation of the L1 norm.

7. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, The pre-training phase, which involves pre-training the initial object detection model, includes: Normalization was performed on the SAR images in the generated domain dataset and the target domain dataset, respectively.

8. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, In the pre-training phase, based on the generator domain dataset and the target domain dataset, and using object detection loss, representative feature enhancement loss, and decoupling domain matching loss, the initial object detection model is pre-trained to obtain a pre-trained object detection model, including: Optical images from the generated domain dataset are input into the initial target detection model, and the target detection loss is calculated based on the detection results output by the initial target detection model and the corresponding labels. Calculate the corresponding foreground target mask based on the labels corresponding to the optical images in the generated domain dataset and the labels corresponding to the SAR images in the target domain dataset; The optical images from the generated domain dataset and the SAR images from the target domain dataset are input into the backbone network of the target detection model to obtain the corresponding optical features and SAR features. The optical features and SAR features are input into the feature discriminator to obtain the corresponding optical foreground target prediction probability and SAR foreground target prediction probability. Calculate the corresponding representative feature enhancement loss based on the foreground target mask, optical foreground target prediction probability, and SAR foreground target prediction probability; The optical features and SAR features are simultaneously input into two different domain classifiers for the foreground and background, respectively, to obtain the corresponding domain classification probabilities. Calculate the corresponding decoupling domain matching loss based on the foreground target mask and the domain classification probability; The initial object detection model is pre-trained using object detection loss, representative feature enhancement loss, and decoupling domain matching loss to obtain a pre-trained object detection model.

9. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 8, characterized in that, The formula for calculating the representative feature enhancement loss is as follows: ； ； in, This represents the loss in representative feature enhancement of an optical image. These represent the pixel width and height of the input feature, respectively. Represents the number of pixels in the optical image in the generated domain dataset. Foreground target mask, Indicates a feature discriminator, Represents the number of pixels in the optical image in the generated domain dataset. The optical features obtained from the backbone network of the target detection model express The predicted probability of the foreground target output after passing through the feature discriminator. This represents the representative feature enhancement loss of the SAR image. Represents the pixels of the SAR image in the target domain dataset. Foreground target mask, Represents the pixels of the SAR image in the target domain dataset. SAR features obtained from the backbone network in the target detection model express The predicted probability of the foreground target output after passing through the feature discriminator; After decoupling the domain matching loss for the foreground domain classifier The calculation is expressed by the formula: ； in, Indicates the number of features of the optical foreground target. Indicates the number of SAR foreground target features. This represents a domain classifier for the foreground. express The domain classification probability output by the domain classifier applied to the foreground. express The domain classification probability output after passing through a domain classifier for the foreground; Decoupling the domain matching loss from the background domain classifier The calculation is expressed by the formula: ； in, Indicates the number of optical background target features. Indicates the number of SAR background target features. This represents a domain classifier for the context. express The domain classification probability output by the domain classifier applied to the background. express The domain classification probability output after passing through a domain classifier for the background.

10. The method for aircraft detection in small-sample SAR images based on dual-level enhancement adaptation according to claim 1, characterized in that, Before training the pre-trained object detection model during the fine-tuning phase, the following steps are included: Normalize the SAR images in the target domain dataset.

Images