A low, slow and small moving target sample generation method based on a degradation model

Abstract

The application discloses a low, slow and small moving target sample generation method based on a degradation model and relates to the technical field of image processing, and specifically comprises the following steps: preparing a data set; acquiring infrared target and background samples; constructing an image degradation model; constructing a target pseudo sample generation model; pseudo data set comparison detection experiment; and pseudo data set sample comparison experiment. The application generates high-quality training data by simulating various degradation phenomena (such as motion blur, optical noise, illumination change and the like) in a real environment, constructs a direction angle controllable generative adversarial network, gradually learns a low-dimensional manifold of a target under continuous direction angles, generates target images of specific direction angles, and is optimized through adversarial training of a loss function, so that high-quality fake low, slow and small moving target images are finally generated; different degradation modes are constructed for the target and the infrared background, the target motion characteristics and the background imaging mechanism are comprehensively considered, and training samples consistent with actual scenes are generated.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for generating low-slow-moving, small-scale target samples based on a degradation model. Background Technology

[0002] Sample generation is a key research topic in computer vision, focusing on creating new data samples to enhance the training effect and generalization ability of models. With the rapid development of military technology and the increasing security demands, moving small target detection has gradually become an important research direction in the field of intelligent perception. In battlefield environments, real-time monitoring of targets such as high-speed flying drones, artillery shells, or missiles is crucial for precision strikes and threat assessment. However, acquiring training data for small targets in real-world scenarios is costly, and due to the special nature of military missions, it may also involve certain security risks. This makes data scarcity a major bottleneck limiting the performance of target detection models. To address these issues, sample generation techniques are being used to expand the amount of data and improve the training effect of detection models. Among these, sample generation methods based on the principle of optical degradation have gradually become a research hotspot. This method generates realistic training samples by simulating imaging degradation phenomena in real environments (such as motion blur, noise interference, and illumination changes), effectively compensating for the lack of real data and providing more diverse training data. Through these generated samples, target detection models can better adapt to complex scenarios and significantly enhance their generalization ability.

[0003] Chinese patent publication number "202310531817.0", entitled "A Method for Generating Infrared Target Samples", aims to use modeling software to assign material values ​​and map textures to different structures of the target object, establishing a three-dimensional geometric model of the target object. It constructs a background model, a temperature field calculation module, an atmospheric radiation transfer calculation module, an infrared radiation characteristic calculation module, and an infrared detector effect calculation module to generate the final infrared target sample, thereby establishing an infrared detector model. However, this method still has some shortcomings in practical applications. The generated sample images may have missing details or be blurry in some cases, especially at target edges and in areas with complex textures, which may affect the performance of subsequent infrared target recognition tasks. Therefore, how to further optimize this method to improve its robustness and computational efficiency in complex environments is an important problem that this invention urgently needs to solve. Summary of the Invention

[0004] The technical solution of this invention to solve the above-mentioned technical problems is to provide a method for generating low-speed, small-motion target samples based on a degradation model, comprising the following steps:

[0005] S1. Prepare the dataset: Nine datasets are used, of which datasets 1, 2, 3 and 6 are used for generating pseudo-samples; datasets 4 and 5 are used for comparative experiments to train the model; datasets 5, 7, 8 and 9 are used for comparative experiments with the generated pseudo-sample datasets.

[0006] S2. Obtain infrared target and background samples: Obtain positive samples of "low, slow, and small" moving targets through dataset six, and generate pseudo samples of "low, slow, and small" moving targets through a generative adversarial network with controllable azimuth angle; generate background pseudo samples by mixing dataset one, dataset two, and dataset three;

[0007] S3. Construct an image degradation model: This includes the target degradation process and the background degradation process. Input the target pseudo-samples and background pseudo-samples generated in step S2 into the model to obtain the degraded sample library.

[0008] S4. Construct a target pseudo-sample generation model: Input the degraded sample library from step S3 into the model to generate infrared weak target pseudo-sample images;

[0009] S5. Pseudo-dataset comparison and detection experiment: Multiple algorithm models are used to detect the generated pseudo-sample data;

[0010] S6. Pseudo-dataset sample comparison experiment: The generated dataset is compared with four classic infrared weak target pseudo-sample datasets to verify the effectiveness of the generation method.

[0011] Furthermore, in S1:

[0012] Dataset 1 is the MDvsFA dataset, Dataset 2 is the SIRST-Aug dataset, and Dataset 3 is the SIRST dataset;

[0013] Dataset 4 is the NUAA-SIRST dataset, and dataset 5 is the IRSD-1k dataset;

[0014] Dataset 6 is a self-made infrared dataset containing weak targets;

[0015] Dataset 7 is the NUDT-SIRST dataset, Dataset 8 is the NUST-SIRST dataset, and Dataset 9 is the IRDST-simulation dataset.

[0016] Furthermore, the azimuth-controllable generative adversarial network in S2 includes:

[0017] Generator: Used to extract and fuse target features from input images with adjacent azimuth angles, generating pseudo-samples between the two azimuth angles;

[0018] Similarity discriminator: used to evaluate the similarity between generated images and real images;

[0019] Azimuth predictor: used to predict the azimuth angle of the generated image;

[0020] Loss functions: The generator uses adversarial loss function, content loss function, perceptual loss function, synthesis loss function and overall change loss function; the similarity discriminator uses adversarial loss function.

[0021] Furthermore, the image degradation model in S3 includes:

[0022] Target degradation: Degradation is performed using isotropic / anisotropic Gaussian blur kernels and multi-directional motion blur functions;

[0023] Background degradation: Degradation is performed using a noise degradation module, which includes Gaussian noise, Poisson noise, JPEG compression noise, and sensor noise.

[0024] Furthermore, the pseudo-sample generation model in S4 adopts:

[0025] Quantity constraint: Controls the number of targets added to the background image within a preset range;

[0026] Position constraints: Ensure that the target is positioned reasonably within the background image.

[0027] Furthermore, the detection algorithm model used in S5 includes:

[0028] YOLOv10s, TLLCM, CNN, Faster R-CNN, and GAN were used, with model training based on datasets three and four.

[0029] Furthermore, the control experiment in S6 includes:

[0030] The generated pseudo-sample dataset is compared with IRSTD-1k dataset 5, NUDT-SIRST dataset 7, NUST-SIRST dataset 8, and IRSTD-simulation dataset 9.

[0031] The detection and validation were performed using a YOLOv10s model trained on datasets four and five.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) This invention proposes a generative adversarial network with controllable azimuth angle, aiming to solve the problem of insufficient image data for moving small targets. This network can generate accurate images of moving small targets with azimuth angles between the azimuth angles of two given images, which helps to enhance the identification and analysis of moving small targets in the case of scarce data.

[0034] (2) This invention proposes a composite loss function consisting of adversarial loss, content loss, perceptual loss, synthesis loss and overall change loss, which makes the generated pseudo sample dataset images more realistic;

[0035] (3) This invention designs a degradation method for "low, slow, and small" moving targets. In actual imaging systems, target degradation is often accompanied by motion blur and noise interference, which significantly affect the observation quality and visual characteristics of the target. By introducing a Gaussian motion blur kernel and noise modeling, the local degradation process of the target in a dynamic environment can be effectively simulated, providing more realistic input data for the training of the detection model and the image restoration algorithm.

[0036] (4) This invention designs a method for degrading infrared background images. Background degradation in infrared images is a common problem in infrared imaging systems, which significantly affects image quality and visual effects, and further reduces the performance of target detection algorithms. This invention simulates noise interference in real imaging environments by targeting Gaussian noise, Poisson noise, JPEG compression with different compression qualities, and sensor noise generated by reverse ISP at different noise levels.

[0037] (5) This invention conducts extensive comparative detection experiments on pseudo-datasets to verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model. Experimental results show that the samples generated by this invention perform well in terms of the accuracy and diversity of target features, and can effectively simulate the characteristics of real "low, slow, and small" moving targets. Furthermore, through comparative detection with other existing sample generation methods, the samples generated by this invention show a more significant performance improvement in subsequent target detection tasks, verifying its effectiveness in improving detection accuracy and robustness.

[0038] (6) This invention conducts extensive comparative experiments with pseudo-dataset samples to verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model. In the experiments, this invention first utilizes the proposed sample generation method, combined with various degradation models, to generate a series of pseudo-datasets with different degrees of degradation and characteristics. These pseudo-datasets cover various complex scenarios that "low, slow, and small" moving targets may encounter in practical applications, including different environmental noise levels, changes in lighting conditions, and differences in target speed and direction. Subsequently, this invention conducts a detailed comparative analysis of these generated pseudo-datasets with existing real datasets and other pseudo-datasets generated based on traditional methods. The comparison includes aspects such as sample feature similarity, target detection performance improvement, model training efficiency, and adaptability under different complex scenarios. Through these comparative experiments, this invention not only verifies the effectiveness of the sample generation method based on the degradation model in generating high-quality, highly realistic samples, but also demonstrates its significant advantages in improving the performance of the target detection system, enhancing the model's generalization ability, and reducing dependence on real data.

[0039] The technical solution of this invention combines an azimuth-controllable GAN with a blind super-resolution degradation model to solve the problems of fuzzy and insufficient diversity in sample generation by traditional methods. It simultaneously covers target motion blur and background noise interference, achieving full-link degradation simulation for the first time in the generation of weak infrared targets. By imposing quantity and position constraints, it improves the rationality of samples, avoiding the "irrational sample" problem common in adversarial networks. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the steps in the method for generating low-slow-moving, small-scale target samples based on a degradation model according to the present invention.

[0042] Figure 2 This is a schematic diagram of the orientation-controllable small target image generation network of the present invention;

[0043] Figure 3 This is a schematic diagram of the generator, discriminator, and predictor of the present invention;

[0044] Figure 4 A diagram showing pseudo-sample data generated by the motion target sample generation method of the present invention and a comparison of pseudo-sample data with other datasets;

[0045] Figure 5 This is a schematic diagram of the detection results of the pseudo-dataset comparison detection experiment of the present invention;

[0046] Figure 6 This is a schematic diagram illustrating the experimental results of comparing samples from the pseudo-dataset described in this paper. Detailed Implementation

[0047] This invention proposes a method for generating low-speed, small-motion target samples based on a degradation model. The aim is to construct a high-order degradation model that comprehensively considers the target motion characteristics and the imaging mechanism of the optical imaging system to generate training samples that match the actual scene.

[0048] The method for generating low-speed, small-motion target samples based on a degradation model proposed in this invention will be described below in specific embodiments:

[0049] Example 1:

[0050] In the technical solution of this embodiment, such as Figure 1 As shown, a method for generating low-slow, small-motion target samples based on a degradation model includes the following steps:

[0051] S1. Prepare the dataset: Nine datasets are used, of which datasets 1, 2, 3 and 6 are used for generating pseudo-samples; datasets 4 and 5 are used for comparative experiments to train the model; datasets 5, 7, 8 and 9 are used for comparative experiments with the generated pseudo-sample datasets.

[0052] Specifically, dataset 1 is the MDvsFA dataset, dataset 2 is the SIRST-Aug dataset, and dataset 3 is the SIRST dataset;

[0053] Dataset 4 is the NUAA-SIRST dataset, and dataset 5 is the IRSD-1k dataset;

[0054] Dataset 6 is a self-made infrared dataset containing weak targets;

[0055] Dataset 7 is the NUDT-SIRST dataset, Dataset 8 is the NUST-SIRST dataset, and Dataset 9 is the IRDST-simulation dataset.

[0056] S2. Obtain infrared target and background samples: Obtain positive samples of "low, slow, and small" moving targets through dataset six, and generate pseudo samples of "low, slow, and small" moving targets through a generative adversarial network with controllable azimuth angle; generate background pseudo samples by mixing dataset one, dataset two, and dataset three;

[0057] Specifically, the azimuth-controllable generative adversarial network (GAN) extracts and fuses target features from two input images with adjacent azimuth angles by designing a specific architecture and input format, and generates a fake image with azimuth angles between the two. The network includes a generator, a similarity discriminator, and an azimuth predictor, which are optimized through adversarial training using a loss function to ultimately generate high-quality fake images.

[0058] Generator: Used to extract and fuse target features from input images with adjacent azimuth angles, generating pseudo-samples between the two azimuth angles;

[0059] Similarity discriminator: used to evaluate the similarity between generated images and real images;

[0060] Azimuth predictor: used to predict the azimuth angle of the generated image;

[0061] Loss functions: The generator uses adversarial loss function, content loss function, perceptual loss function, synthesis loss function and overall change loss function; the similarity discriminator uses adversarial loss function.

[0062] S3. Construct an image degradation model: This includes the target degradation process and the background degradation process. Input the target pseudo-samples and background pseudo-samples generated in step S2 into the model to obtain the degraded sample library.

[0063] Specifically, image degradation models include:

[0064] Target degradation: Degradation is performed using isotropic / anisotropic Gaussian blur kernels and multi-directional motion blur functions; the Gaussian blur kernel is composed of anisotropic Gaussian blur functions and multi-directional motion blur functions, providing theoretical support for the realistic generation of small sample targets in complex environments;

[0065] Background degradation: Degradation is performed using a noise degradation module, which includes Gaussian noise, Poisson noise, JPEG compression noise, and sensor noise. Noise interference in the real imaging environment is simulated by adjusting the noise level and compression quality factor.

[0066] S4. Construct a target pseudo-sample generation model: Input the degraded sample library from step S3 into the model to generate infrared weak target pseudo-sample images;

[0067] Specifically, the pseudo-sample generation model employs:

[0068] Quantity constraint: Control the number of targets added to the background image within a preset range to ensure that the number of targets added to the overall infrared background image is within a reasonable range;

[0069] Position constraints: Ensure that the target's position and layout in the background image are reasonable, and constrain the target's overall position and layout in the background image to be more appropriate;

[0070] S5. Pseudo-dataset comparison and detection experiment: Multiple algorithm models are used to detect the generated pseudo-sample data;

[0071] The detection algorithm models used include YOLOv10s, TLLCM, CNN, Faster R-CNN and GAN. The model training is based on datasets three and four to verify the effectiveness of the "low, slow and small" moving target sample generation method based on the degradation model.

[0072] S6. Pseudo-dataset sample comparison experiment: The generated dataset is compared with four classic infrared weak target pseudo-sample datasets to verify the effectiveness of the generation method.

[0073] The control experiments included: comparing the generated pseudo-sample dataset with IRSTD-1k dataset 5, NUDT-SIRST dataset 7, NUST-SIRST dataset 8, and IRSTD-simulation dataset 9; and using a YOLOv10s model trained on datasets 4 and 5 for detection and verification.

[0074] Example 2:

[0075] A method for generating low-speed, small-motion target samples based on a degradation model includes the following steps:

[0076] S1. Dataset Preparation: A multi-source, heterogeneous infrared weak target dataset system was adopted, comprehensively covering the detection needs of different scenarios and characteristics. Among them, the MDvsFA dataset 1 contains moving target sequences against complex backgrounds, particularly suitable for verifying the robustness of the algorithm in dynamic scenes. The SIRST-Aug dataset 2 and SIRST dataset 3 provide rich single-frame infrared images, including stationary targets against various typical ground cover backgrounds, with a wide range of target sizes. The NUAA-SIRST dataset 4, constructed by Nanjing University of Aeronautics and Astronautics, focuses on collecting aerial weak target scenes and has accurate radiometric characterization. The IRSD-1k dataset 5 targets long-distance observation scenarios, where the target signal-to-noise ratio is generally below 3dB, posing a severe challenge to the algorithm's sensitivity. To compensate for the shortcomings of existing datasets, a self-built infrared weak target dataset 6 was also introduced, with "low, slow, and small" moving targets ranging in size from 3×3 pixels to 15×15 pixels. The NUDT-SIRST dataset VII and NUST-SIRST dataset VIII are from the National University of Defense Technology and Nanjing University of Science and Technology, respectively. The former focuses on military application scenarios, while the latter contains a large number of target samples under simulated tactical environments. The IRDST-simulation dataset IX is generated by a high-fidelity infrared physical simulation system, which can accurately control the target's motion trajectory, radiation characteristics, and background interference, providing an ideal platform for evaluating the algorithm's performance under extreme conditions.

[0077] S2. Obtain infrared target and background samples: Extract positive sample data of "low, slow, and small" moving targets from dataset six in step S1. According to manifold learning theory, high-dimensional data is often mapped from low-dimensional manifolds, and the target distribution in images of small moving targets is also like this. A network for generating images of small moving targets with controllable orientation, such as... Figure 2 As shown, the network includes feature extraction, residual blocks, and mapping blocks. First, two input images with adjacent azimuth angles are processed by their respective feature extraction modules. These features are then fused in the residual block, which utilizes skip connections to help preserve and adjust the features. Finally, the mapping block transforms the fused features into a new "low, slow, small" infrared small target image. The generated image is fed into a similarity discriminator and an azimuth predictor to evaluate its similarity to the real image and predict its azimuth angle, thus ensuring the quality of the generated image and the accuracy of the azimuth angle. This process is optimized through adversarial training, ultimately generating high-quality fake images. This network achieves a more stable training process than traditional GANs and generates accurate moving small target images with controllable azimuth angles.

[0078] The general process of a position-controllable moving small target image generation network progressively learning a low-dimensional manifold can be described as follows: Given two moving small target images I1 and I2 with adjacent azimuth angles θ1 and θ2 as input to the generator G, G will generate images with azimuth angles θ1 < θ2. g Image of a pseudo-moving small target <θ2 I g However, in the initial state, the false moving small target image I g With θ g The true moving small target image Ir has almost no identical information. Then, the fake I... g Kazuma I r The data is fed into the discriminator and predictor. The similarity discriminator determines whether each image is real or fake, while the orientation predictor predicts the azimuth angles of the two images. Through adversarial training, the generator and discriminator are optimized. Ultimately, the network can generate images of small moving targets with precise features and controllable azimuth angles.

[0079] like Figure 3 The diagram shows the network structure consisting of a generator, discriminator, and predictor. This network employs a unique topology, incorporating a skip connection design with residual blocks to address the vanishing gradient problem in deep networks, and utilizing the generator to generate data for specific tasks. The orientation predictor and similarity discriminator modules flexibly capture spatial features through deformable convolutions, while multi-scale feature extraction paths process input information at different scales in parallel. Furthermore, the network extensively uses the Leaky ReLU activation function to prevent neuron "death." These design features collectively optimize the network's adaptability and performance for complex tasks.

[0080] The network uses two parallel input blocks B pi1 and B pi2 Extract azimuth information, information fusion block B if Fusion features, mapping block B m Generate images. Images I1 and I2 are input into B respectively. pi1 and B pi2 Learning the target features of adjacent azimuth angles can be defined as:

[0081] l pi1 =B pi1 (I1)

[0082] l pi2 =B pi2 (I2)

[0083] Among them, l pi1 and l pi2 This represents the target features of the two inputs at different azimuth angles. Then, information fusion block B... if By mapping l pi1 and l pi2 Information between them, adaptively acquiring and retaining azimuth angle θ1<θ g The target information of a moving small target image with a value <θ2 can be defined as:

[0084] R interp =B if (l pi1 ,l pi2 ),

[0085] Among them, R interp Indicates azimuth angle θ1 < θ g Mapping information of the moving small target image with a value <θ2. Finally, combining the target information, mapping block B. m Information can be mapped from a high dimension to a two-dimensional image space to generate a final, accurate image of a small moving target.

[0086] To preserve information, residual blocks and batch normalization were used in all three blocks of the network. Next, a similarity discriminator and an azimuth predictor were designed. Both of them used stride convolution instead of pooling layers, allowing the network to learn its own spatial downsampling. Leaked modified linear units (LReLU) activation were used in all layers, while batch normalization was adopted and all dense layers were removed to accommodate a deeper architecture. For the similarity discriminator, the weights were pruned after flattening the last layer to optimize the EM distance; for the azimuth predictor, deformable convolution was used to capture azimuth changes, and vector multiplication was performed after flattening the last layer

[35] .

[0087] As mentioned above, the total value function of the proposed orientation-controllable moving small target image generation network can be expressed as:

[0088]

[0089] Where E represents the expectation operator, Eu represents the Euclidean distance, and I r Let I1 and I2 represent the two input real moving small target images to the discriminator and predictor, respectively, and let Pdata(I) represent the distribution of the real moving small target images. Let θ represent the input real moving small target images to the generator. g This represents the azimuth angle of the real moving small target image as input to the predictor. The first two terms of F1 are to ensure that the generated moving small target image can be accurately distinguished by the discriminator. The third term of F1 is to enable the generator to generate an image with the expected azimuth angle. As for F2, it is the azimuth angle loss of the azimuth angle predictor, which aims to minimize the azimuth angle distance between the real moving small target image and the generated moving small target image.

[0090] Therefore, the loss of the similarity discriminator can be expressed as:

[0091]

[0092] For azimuth predictors, a mean squared error is introduced. The loss of the azimuth predictor can be expressed by the following formula:

[0093]

[0094] Where I r θ represents the input image of a real moving small target to the discriminator and predictor. g This represents the azimuth angle of the real moving small target image that serves as the input to the predictor.

[0095] As for the generator's loss function, the Wasserstein-1 distance is used instead of the Jensen-Shannon divergence in traditional GANs. It can be defined as:

[0096]

[0097] As the proposed orientation-controlled moving small target image generation network is trained, the generator, discriminator, and predictor are updated alternately for optimization. Therefore, when the discriminator and predictor are able to more accurately identify the image and predict the azimuth, the generator can make the generated moving small target image closer to the real image in terms of similarity and azimuth.

[0098] In terms of background data preparation, using datasets 1, 2 and 3 from step S1, these background data are fused together using an intelligent hybridization algorithm to construct a background sample library with rich scene variations.

[0099] S3. Constructing an Image Degradation Model: The image degradation model includes target degradation and background degradation. The pseudo-samples of "low, slow, and small" moving targets and backgrounds obtained in step S2 are input into the image degradation model to obtain a degraded sample library. Motion blurring simulation combines uniform linear motion and random jitter modes, achieving different intensities of blurring through an adjustable blur kernel function (σ∈[0.5,3.0]). Resolution degradation employs an adaptive downsampling mechanism, randomly selecting a downsampling factor (2× to 4×) and using a bicubic interpolation algorithm while maintaining target recognizability. Specifically, the model introduces a random scheduling mechanism for the degradation process, dynamically adjusting the application order and parameter combinations of each degradation module through a Markov decision process to ensure that the generated degradation samples conform to physical laws and possess sufficient diversity. The model takes the pseudo-samples generated in stage S2 as input, and after the above degradation processing, outputs a sample library containing different degradation features, where each sample is accompanied by complete degradation parameter labels, providing reliable benchmark data for subsequent algorithm training. This modeling approach not only covers common imaging degradation factors, but also preserves the interpretability of the degradation process through parametric design, which is beneficial for model optimization and performance analysis for specific scenarios.

[0100] S4. Constructing a target pseudo-sample generation model: Input the degraded sample library obtained in S3 into the pseudo-sample generation model to generate infrared weak target pseudo-sample images. A pseudo-sample generation system for infrared weak targets based on deep generative adversarial networks was designed. Figure 4The generated pseudo-sample data is shown. In the generator design, the U-Net++ architecture is used as the backbone network, embedding deformable convolutional modules to simulate the deformation characteristics of the target under different observation angles. Simultaneously, a physical constraint layer based on infrared radiation characteristics is introduced to ensure that the generated targets have reasonable temperature distribution and radiation characteristics. The discriminator adopts a multi-branch structure, distinguishing between real and fake targets from three dimensions: local texture details, global radiation distribution, and motion blur features. The local discrimination branch is specifically optimized for the edge sharpness and signal-to-noise ratio characteristics of small targets. A progressive learning strategy is used during training, first learning the overall motion trajectory of the target in a large-scale feature space, and then gradually focusing on small-scale features to extract the fine texture of the target. To improve the diversity of generated samples, the system integrates a conditional variational autoencoder, which continuously adjusts key parameters such as target size, motion speed, and radiation intensity by introducing random perturbation vectors into the latent space. In particular, the generation model also incorporates prior knowledge of the infrared physical imaging process, ensuring that the generated pseudo-targets conform to the Stefan-Boltzmann law through constraints from the thermal radiation equation. The final output pseudo-sample images not only possess realistic infrared characteristics but also include complete metadata tags, such as the target's equivalent blackbody temperature, motion parameters, and background complexity, providing rich supervisory signals for the training and evaluation of subsequent detection algorithms. This generation system, through end-to-end training, achieves intelligent conversion from degraded samples to high-quality pseudo-samples, effectively solving the problem of obtaining realistic infrared weak target samples. Figure 4 This diagram illustrates a comparison between the generated pseudo-samples, those generated by the CycleGAN method, and those generated by the DCGAN method. The method presented in this embodiment demonstrates a significant advantage in generating high-quality, highly consistent pseudo-datasets, particularly in generating datasets containing a large number of "low, slow, and small" moving targets. These datasets are crucial for simulating and testing complex air traffic monitoring systems, drone defense systems, and various automated monitoring systems. Compared to the CycleGAN and DCGAN generation methods, the method of this invention outperforms the method in terms of the quality, stability, and consistency of the generated data. The images generated by this invention are closer to the original images in terms of detail and grayscale distribution, and maintain high accuracy in target scale and spatial structure. In contrast, CycleGAN and DCGAN exhibit significant shortcomings in generating complex scenes, particularly in grayscale distribution, color shift, and spatial structure.

[0101] S5. Pseudo-Dataset Comparison and Detection Experiment: To verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model, YOLOv10s, TLLCM, CNN, Faster R-CNN, and GAN algorithms were used to detect the generated small target sample data. The experiment was conducted on a high-performance computing platform equipped with an NVIDIA Quadro RTX 4090 (32GB VRAM), using the Ubuntu 18.04 operating system environment and 8GB of RAM to ensure stable system operation. A complete GPU-accelerated computing environment was constructed, using the CUDA 12.0 parallel computing architecture and the PyTorch deep learning framework for model training and validation. In terms of the software environment, the flexibility and scalability of the PyTorch framework created favorable conditions for rapid model iteration and experimental validation, while the optimized computing power of CUDA 12.0 significantly improved model training efficiency, ensuring that the complete algorithm validation process could be completed within a reasonable time. Figure 5 This is a qualitative result.

[0102] S6. Pseudo-Dataset Sample Comparison Experiment: To further verify the effectiveness of the degradation model-based method for generating "low, slow, and small" moving target samples, a comparison experiment was conducted using the generated dataset and two classic infrared weak target pseudo-sample datasets. The experiment used datasets four and five to train the model on a YOLOv10s network architecture, and a unified evaluation metric was used to compare the model's performance on real infrared datasets. The evaluation metrics used were IoU and false alarm rate as pixel-level evaluation metrics, and detection probability was used to evaluate target-level performance. Furthermore, a comparison experiment was conducted using the generated pseudo-dataset with IRSTD-1k dataset five, NUDT-SIRST dataset seven, NUST-SIRST dataset eight, and IRSTD-simulation dataset nine. Quantitative analysis further verified the effectiveness of the degradation model-based method for generating "low, slow, and small" moving target samples. Figure 6 This is a quantitative result.

[0103] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating low-speed, small-motion target samples based on a degradation model, characterized in that, Includes the following steps: S1. Prepare the dataset: Nine datasets are used, of which datasets 1, 2, 3 and 6 are used for generating pseudo-samples; datasets 4 and 5 are used for comparative experiments to train the model; datasets 5, 7, 8 and 9 are used for comparative experiments with the generated pseudo-sample datasets. S2. Obtain infrared target and background samples: Obtain positive samples of "low, slow, and small" moving targets through dataset six, and generate pseudo samples of "low, slow, and small" moving targets through a generative adversarial network with controllable azimuth angle; generate background pseudo samples by mixing dataset one, dataset two, and dataset three; S3. Construct an image degradation model: This includes the target degradation process and the background degradation process. Input the target pseudo-samples and background pseudo-samples generated in step S2 into the model to obtain the degraded sample library. S4. Construct a target pseudo-sample generation model: Input the degraded sample library from step S3 into the model to generate infrared weak target pseudo-sample images; S5. Pseudo-dataset comparison and detection experiment: Multiple algorithm models are used to detect the generated pseudo-sample data; S6. Pseudo-dataset sample comparison experiment: The generated dataset is compared with four classic infrared weak target pseudo-sample datasets to verify the effectiveness of the generation method; In S1: Dataset 1 is the MDvsFA dataset, Dataset 2 is the SIRST-Aug dataset, and Dataset 3 is the SIRST dataset; Dataset 4 is the NUAA-SIRST dataset, and dataset 5 is the IRSTD-1k dataset; Dataset 6 is a self-made infrared dataset containing weak targets; Dataset 7 is the NUDT-SIRST dataset, Dataset 8 is the NUST-SIRST dataset, and Dataset 9 is the IRDST-simulation dataset.

2. The method for generating low-speed, small-motion target samples based on a degradation model according to claim 1, characterized in that, The azimuth-controllable generative adversarial network in S2 includes: Generator: Used to extract and fuse target features from input images with adjacent azimuth angles, generating pseudo-samples between the two azimuth angles; Similarity discriminator: used to evaluate the similarity between generated images and real images; Azimuth predictor: used to predict the azimuth angle of the generated image; Loss functions: The generator uses adversarial loss function, content loss function, perceptual loss function, synthesis loss function and overall change loss function; the similarity discriminator uses adversarial loss function.

3. The method for generating low-speed, small-motion target samples based on a degradation model according to claim 1, characterized in that, The image degradation model in S3 includes: Target degradation: Degradation is performed using isotropic / anisotropic Gaussian blur kernels and multi-directional motion blur functions; Background degradation: Degradation is performed using a noise degradation module, which includes Gaussian noise, Poisson noise, JPEG compression noise, and sensor noise.

4. The method for generating low-speed, small-motion target samples based on a degradation model according to claim 1, characterized in that, The pseudo-sample generation model in S4 uses: Quantity constraint: Controls the number of targets added to the background image within a preset range; Position constraints: Ensure that the target is positioned reasonably within the background image.

5. The method for generating low-speed, small-motion target samples based on a degradation model according to claim 1, characterized in that, The detection algorithm model used in S5 includes: YOLOv10s, TLLCM, CNN, Faster R-CNN, and GAN were used, with model training based on datasets three and four.

6. The method for generating low-speed, small-motion target samples based on a degradation model according to claim 1, characterized in that, The control experiment in S6 includes: The generated pseudo-sample dataset is compared with IRSTD-1k dataset 5, NUDT-SIRST dataset 7, NUST-SIRST dataset 8, and IRSTD-simulation dataset 9. The detection and validation were performed using a YOLOv10s model trained on datasets four and five.

Images