Low-slow small moving target sample generation method based on degradation model

Abstract

The invention discloses a low-slow small moving target sample generation method based on a degradation model, and relates to the technical field of image processing, and the method specifically comprises the following steps: preparing a data set; obtaining an infrared target and a background sample; constructing an image degradation model; constructing a target pseudo sample generation model; performing a pseudo data set comparison detection experiment; and performing a pseudo data set sample contrast experiment. By simulating various degradation phenomena (such as motion blur, optical noise and illumination variation) in a real environment, high-quality training data is generated, a generative adversarial network with a controllable azimuth angle is constructed, a low-dimensional manifold of a target under a continuous azimuth angle is learned step by step, a target image with a specific azimuth angle is generated, and the target image with the specific azimuth angle is obtained. And through loss function adversarial training optimization, a high-quality false'low-slow-small 'moving target image is finally generated. Different degradation modes are constructed for a target and an infrared background, target motion characteristics and a background imaging mechanism are comprehensively considered, and a training sample conforming to an actual scene is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method for generating low, slow, and small moving target samples based on a degradation model. Background Art

[0002] Sample generation is a key research topic in the field of computer vision, which focuses on creating new data samples to enhance the training effect and generalization ability of models. With the rapid development of military technology and the continuous improvement of security requirements, the detection of small moving targets has gradually become an important research direction in the field of intelligent perception. In the battlefield environment, real-time monitoring of targets such as high-speed flying drones, shells, or missiles is of great significance for precise strikes and threat assessment. However, the cost of obtaining training data for small targets in real scenarios is high, and due to the particularity of military tasks, there may also be certain security risks. This makes data scarcity the main bottleneck restricting the performance of target detection models. To solve the above problems, attempts are made to expand the data volume through sample generation technology to improve the training effect of detection models. Among them, the sample generation method based on the optical degradation principle has gradually become a research hotspot. This method generates realistic training samples by simulating imaging degradation phenomena in the real environment (such as motion blur, noise interference, and illumination changes), thus effectively making up for the deficiency of real data and providing more diverse training data. Through these generated samples, the target detection model can better adapt to complex scenarios and significantly enhance its generalization ability.

[0003] The Chinese patent publication number is "202310531817.0", and the name is "A method for generating infrared target samples", aiming to assign materials and texture maps to different structures of the target object using modeling software to establish a three-dimensional geometric model of the target object. Construct a background model, a temperature field calculation module, an atmospheric radiation transmission calculation module, an infrared radiation characteristic calculation, and an infrared detector effect calculation module to generate the final infrared target samples, thereby establishing an infrared detector model. However, there are still some deficiencies in the actual application of this method. The generated sample images may have problems such as missing details or blurring in some cases, especially in areas with complex target edges and 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 the present invention urgently needs to solve. Summary of the Invention

[0004] The technical solution of the present invention to solve the above technical problems is to provide a method for generating low, slow, and small moving target samples based on a degradation model, including the following steps:

[0005] S1. Prepare the dataset: Use nine datasets, where Dataset 1, Dataset 2, Dataset 3, and Dataset 6 are used for pseudo-sample generation; Dataset 4 and Dataset 5 are used for training the model in the comparative experiment; Dataset 5, Dataset 7, Dataset 8, and Dataset 9 are used for the comparative experiment with the generated pseudo-sample dataset.

[0006] S2. Obtain infrared target and background samples: Obtain positive samples of "low, slow, and small" moving targets through Dataset 6, 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 1, Dataset 2, and Dataset 3.

[0007] S3. Construct an image degradation model: Include the target degradation process and the background degradation process, and input the target pseudo-samples and background pseudo-samples generated in step S2 into this model to obtain a degraded sample library.

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

[0009] S5. Comparative detection experiment on the pseudo-dataset: Use various algorithm models to detect the generated pseudo-sample data.

[0010] S6. Control experiment on the pseudo-dataset samples: Compare the generated dataset with four classic infrared small and weak target pseudo-dataset to verify the effectiveness of the generation method.

[0011] Furthermore, in the above 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 small and 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 generative adversarial network with controllable azimuth angle in the above S2 includes:

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

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

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

[0020] Loss function: The generator adopts adversarial loss function, content loss function, perceptual loss function, synthesis loss function and total variation loss function; The similarity discriminator adopts adversarial loss function.

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

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

[0023] Background degradation: Degradation is carried out using a noise degradation module, including Gaussian noise, Poisson noise, JPEG compression noise and sensor noise.

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

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

[0026] Position constraint: Ensure that the position layout of the targets in the background image is reasonable.

[0027] Furthermore, the detection algorithm models adopted in S5 include:

[0028] YOLOv10s, TLLCM, CNN, Faster R-CNN and GAN, where the model training is based on Dataset Three and Dataset Four.

[0029] Furthermore, the control experiments in S6 include:

[0030] Compare the generated pseudo-sample dataset with IRSTD-1k Dataset Five, NUDT-SIRST Dataset Seven, NUST-SIRST Dataset Eight and IRDST-simulation Dataset Nine;

[0031] Use the YOLOv10s model trained based on Dataset Four and Dataset Five for detection and verification.

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

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

[0034] (2) The present invention proposes a composite loss function composed of adversarial loss, content loss, perceptual loss, synthesis loss, and overall variation loss, making the generated images of the pseudo-sample dataset have higher authenticity;

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

[0036] (4) The present invention designs a degradation for infrared background images. Infrared image background degradation is a common problem in infrared imaging systems, which will significantly affect the image quality and visual effects, and further reduce the performance of target detection algorithms. The present invention simulates the noise interference in a real imaging environment for Gaussian noise, Poisson noise with different noise levels, JPEG compression with different compression qualities, and sensor noise generated by reversing ISP.

[0037] (5) The present invention conducts a large number of pseudo-dataset comparison detection experiments to verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model. The experimental results show that the samples generated by the present invention perform excellently in terms of the accuracy and diversity of target features, and can effectively simulate the characteristics of real "low, slow, and small" moving targets. At the same time, through the comparison detection with other existing sample generation methods, the performance of the samples generated by the method of the present invention is more significantly improved in subsequent target detection tasks, verifying its effectiveness in improving detection accuracy and robustness.

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

[0039] The technical solution of the present invention combines an azimuth - controllable GAN with a blind super - resolution degradation model to solve the problems of blurred and insufficiently diverse samples generated by traditional methods. At the same time, it covers target motion blur and background noise interference, and for the first time realizes full - link degradation simulation in the generation of infrared small and weak targets. By imposing quantity and position constraints, it improves the rationality of samples and avoids the common "unreasonable sample" problem in generative adversarial networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0041] Figure 1 is the flowchart of the steps of the method for generating samples of "low - slow - small" moving targets based on the degradation model according to the present invention;

[0042] Figure 2 is the structural schematic diagram of the azimuth - controllable small - target image generation network according to the present invention;

[0043] Figure 3 is the structural schematic diagram of the generator, discriminator, and predictor according to the present invention;

[0044] Figure 4 is the schematic diagram of the pseudo - sample data generated by the method for generating samples of moving targets according to the present invention and the comparison between the pseudo - sample data and other datasets;

[0045] Figure 5 Schematic diagram of the detection result of the pseudo-dataset contrast detection experiment of the present invention;

[0046] Figure 6 Schematic diagram of the effect of the pseudo-dataset sample contrast experiment described in the present invention. Specific implementation manners

[0047] The present invention proposes a method for generating low, slow, and small moving target samples based on a degradation model, aiming to construct a high-order degradation model, comprehensively considering the target motion characteristics and the imaging mechanism of the optical imaging system, and generating training samples that conform to the actual scenario.

[0048] The method for generating low, slow, and small moving target samples based on the degradation model proposed by the present invention will be described in the following specific embodiments:

[0049] Embodiment 1:

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

[0051] S1. Prepare the dataset: Use nine datasets, where Dataset 1, Dataset 2, Dataset 3, and Dataset 6 are used for generating pseudo-samples; Dataset 4 and Dataset 5 are used for training the model in the contrast experiment; Dataset 5, Dataset 7, Dataset 8, and Dataset 9 are used for the contrast experiment with the generated pseudo-sample dataset;

[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 and small 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 6, 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 1, Dataset 2, and Dataset 3;

[0057] Specifically, the azimuth - controllable generative adversarial network realizes the extraction and fusion of target features from two input images with adjacent azimuths and generates a fake image with an azimuth between the two by designing a specific architecture and input form. The network includes a generator, a similarity discriminator, and an azimuth predictor, and is optimized through adversarial training with a loss function to finally generate high - quality fake images;

[0058] Generator: It is used to extract and fuse target features from input images with adjacent azimuths and generate pseudo - samples with an azimuth between the two;

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

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

[0061] Loss function: The generator uses an adversarial loss function, a content loss function, a perceptual loss function, a synthesis loss function, and a total variation loss function; the similarity discriminator uses an adversarial loss function.

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

[0063] Specifically, the image degradation model includes:

[0064] Target degradation: Isotropic / anisotropic Gaussian blur kernels and multi - directional motion blur functions are used for degradation; the Gaussian blur kernels are composed of anisotropic Gaussian blur functions and multi - directional motion blur functions, providing theoretical support for the generation of realistic small - sample targets in complex environments;

[0065] Background degradation: A noise degradation module is used for degradation, including Gaussian noise, Poisson noise, JPEG compression noise, and sensor noise. By adjusting the noise level and compression quality factor, the noise interference in the real imaging environment is simulated.

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

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

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

[0069] Position constraint: Ensure that the position layout of the targets in the background image is reasonable, and constrain the addition of targets to make the overall position layout of the background image more reasonable;

[0070] S5. Pseudo-dataset comparison detection experiment: Use multiple algorithm models to detect the generated pseudo-sample data;

[0071] The detection algorithm models used include: YOLOv10s, TLLCM, CNN, Faster R-CNN, and GAN. Among them, the model training is based on Dataset III and Dataset IV, and the validation is based on the "low, slow, and small" moving target sample generation method of the degradation model to verify its effectiveness.

[0072] S6. Pseudo-dataset sample control experiment: Compare the generated dataset with four classic infrared weak and small target pseudo-datasets to verify the effectiveness of the generation method.

[0073] The control experiment includes: Comparing the generated pseudo-sample dataset with IRSTD-1k Dataset V, NUDT-SIRST Dataset VII, NUST-SIRST Dataset VIII, and IRDST-simulation Dataset IX; Using the YOLOv10s model trained based on Dataset IV and Dataset V for detection and verification.

[0074] Example 2:

[0075] A method for generating "low, slow, and small" moving target samples based on a degradation model, comprising the following steps:

[0076] S1. Prepare the dataset: Adopt a multi-source heterogeneous infrared weak and small target dataset system, which comprehensively covers the detection requirements of different scenarios and different characteristics. Among them, MDvsFA Dataset I contains moving target sequences in complex backgrounds, which is especially suitable for verifying the robustness of algorithms in dynamic scenarios. SIRST-Aug Dataset II and SIRST Dataset III provide rich single-frame infrared images, containing stationary targets in various typical ground object backgrounds, and the target size distribution is wide. NUAA-SIRST Dataset IV is constructed by Nanjing University of Aeronautics and Astronautics, focusing on collecting aerial weak target scenarios, with accurate radiation characteristic annotations. IRSD-1k Dataset V is for long-distance observation scenarios, where the target signal-to-noise ratio is generally lower than 3dB, posing a severe challenge to the algorithm sensitivity. To make up for the deficiencies of existing datasets, a self-built infrared weak and small target dataset VI is also introduced, and the "low, slow, and small" moving small target sizes range from 3×3 pixels to 15×15 pixels. NUDT-SIRST Dataset VII and NUST-SIRST Dataset VIII come from the National University of Defense Technology and Nanjing University of Science and Technology respectively. The former focuses on military application scenarios, and the latter contains a large number of target samples in simulated tactical environments. IRDST-simulation Dataset IX, which is generated by a high-fidelity infrared physical simulation system, can accurately control the target motion trajectory, radiation characteristics, and background interference, providing an ideal platform for the performance evaluation of algorithms under extreme conditions.

[0077] S2. Obtain infrared target and background samples: Extract the positive sample data of "low, slow, and small" moving targets from the dataset six in step S1. According to the manifold learning theory, high-dimensional data is often mapped from low-dimensional manifolds, and the same is true for the target distribution of small moving target images. A moving small target image generation network with controllable azimuth, such as Figure 2 shown, which includes feature extraction, residual blocks, and mapping blocks. First, two input images with adjacent azimuth angles obtain features through their respective feature extraction modules, and then these features are fused in the residual blocks. The residual blocks use skip connections to help retain and adjust features. Finally, the mapping blocks convert the fused features into new "low, slow, and small" infrared small target images. The generated images are fed into a similarity discriminator and an azimuth predictor to evaluate their similarity to the real images and predict their azimuth angles, thus ensuring the quality of the generated images and the accuracy of the azimuth angles. This process is optimized through adversarial training, and finally high-quality fake images are generated. This network can obtain a more stable training process than traditional GANs and generate accurate moving small target images with controllable azimuth angles.

[0078] The general process of the moving small target image generation network with controllable azimuth gradually learning the 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 the input of the generator G, the fake moving small target image I g <θ2 generated by G with azimuth angle θ1 < θ g . However, in the initial state, the fake moving small target image I g has almost no information in common with the real moving small target image Ir with θ g . Then, the fake I g and the real I r are input into the discriminator and the predictor. The similarity discriminator will determine whether each image is real or fake, and the azimuth predictor will predict the azimuth angles of the two images. Through adversarial training, the generator and the discriminator are optimized. Finally, the network can generate moving small target images with accurate features and controllable azimuth angles.

[0079] As Figure 3 shown, it is the network structure of the generator, discriminator, and predictor. This network adopts a special topological structure, combines the skip connection design of the residual blocks to solve the gradient vanishing problem of deep networks, and uses the generator to implement data generation for specific tasks. The azimuth predictor and similarity discriminator modules flexibly capture spatial features through deformable convolutions, while the multi-scale feature extraction path processes input information at different scales in parallel. In addition, the network widely uses the Leaky ReLU activation function to avoid neuron "death", and these designs together optimize the adaptability and performance of the network for complex tasks.

[0080] The network extracts azimuth information through two parallel input blocks B pi1 and B pi2 and fuses features in the information fusion block B if and generates an image in the mapping block B m Images I1 and I2 are respectively input into B pi1 and B pi2 to learn the target features of adjacent azimuth angles, which can be defined as:

[0081] l pi1 = B pi1 (I1)

[0082] l pi2 = B pi2 (I2)

[0083] where l pi1 and l pi2 represent the target features of the two inputs at different azimuth angles. Then, the information fusion block B if adaptsively obtains and retains the target information of the moving small target image with azimuth angle θ1 < θ pi1 < θ2 by mapping the information between l pi2 and l g , which can be defined as:

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

[0085] where R interp represents the mapping information of the moving small target image with azimuth angle θ1 < θ g < θ2. Finally, combining the target information, the mapping block B m can map the information from high dimension to two-dimensional image space to generate the final accurate moving small target image.

[0086] To retain information, residual blocks and batch normalization are used in all three blocks of the network. Secondly, a similarity discriminator and an azimuth predictor are designed. They both use stride convolution instead of the pooling layer, allowing the network to learn its own spatial downsampling, and use leaky rectified linear unit (LReLU) activation in all layers, while adopting batch normalization and removing all dense layers to adapt to deeper architectures. For the similarity discriminator, weight clipping is performed after flattening the last layer to optimize the EM distance; for the azimuth predictor, deformable convolution is used to capture azimuth changes, and vector multiplication is performed after flattening the last layer

[35] .

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

[0088]

[0089] Among them, E represents the expectation operator, Eu represents the Euclidean distance, and I r represents the input real moving small target image of the discriminator and the predictor, I1 and I2 represent the two input real moving small target images of the generator, Pdata(I) represents the distribution of the real moving small target image, and θ g represents the azimuth angle of the input real moving small target image of 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 make the generator generate an image with the expected azimuth angle. As for F2, it is the azimuth loss of the azimuth predictor, aiming to minimize the azimuth 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 the azimuth predictor, the mean square error is proposed. The loss of the azimuth predictor can be expressed by the following formula:

[0093]

[0094] where I r represents the input real moving small target image of the discriminator and the predictor, and θ g represents the azimuth angle of the input real moving small target image of the predictor.

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

[0096]

[0097] With the training of the proposed azimuth controllable moving small target image generation network, the generator, discriminator, and predictor are alternately updated for optimization. Therefore, when the discriminator and predictor can more accurately identify the image and predict the azimuth angle, the generator can make the generated moving small target image closer to the real image in terms of similarity and azimuth angle.

[0098] In terms of background data preparation, through Dataset 1, Dataset 2, and Dataset 3 in step S1, these background data are fused by the intelligent hybrid algorithm to construct a background sample library with rich scene variations.

[0099] S3. Construct an image degradation model: The image degradation model includes target degradation and background degradation; the "low, slow, and small" moving target pseudo-samples and background pseudo-samples obtained in step S2 are input into the image degradation model to obtain a degraded sample library. Among them, the motion blur simulation combines two modes of uniform linear motion and random jitter, and different intensity blur effects are achieved through an adjustable blur kernel function (σ ∈ [0.5, 3.0]); the resolution degradation adopts an adaptive downsampling mechanism. On the premise of maintaining the recognizability of the target, the downsampling factor (2× to 4×) is randomly selected and combined with the bicubic interpolation algorithm. In particular, the model introduces a random scheduling mechanism for the degradation process, and dynamically adjusts the application order and parameter combination of each degradation module through a Markov decision process to ensure that the generated degraded samples not only conform to physical laws but also have sufficient diversity. This model takes the pseudo-samples generated in the S2 stage as input, and after the above degradation process, outputs a sample library containing different degradation characteristics. Each sample is attached with a complete degradation parameter label, providing reliable benchmark data for subsequent algorithm training. This modeling method not only covers common imaging degradation factors but also retains the interpretability of the degradation process through parametric design, which is beneficial for model tuning and performance analysis for specific scenarios.

[0100] S4. Construct a target pseudo-sample generation model: The degraded sample library obtained in S3 is input into the pseudo-sample generation model to generate infrared small and weak target pseudo-sample images, and an infrared small and weak target pseudo-sample generation system based on a deep generative adversarial network is designed. Figure 4The generated partial pseudo-sample data is shown. In the generator design, the U-Net++ architecture is adopted as the backbone network, and a deformable convolution module is embedded to simulate the deformation characteristics of the target under different observation angles. At the same time, a physical constraint layer based on infrared radiation characteristics is introduced to ensure that the generated target has a reasonable temperature distribution and radiation characteristics. The discriminator adopts a multi-branch structure to distinguish the authenticity from three dimensions: local texture details, global radiation distribution, and motion blur characteristics. Among them, the local discrimination branch is specifically optimized for the edge sharpness and signal-to-noise ratio characteristics of small and weak targets. During the training process, a progressive learning strategy is adopted. First, the overall motion trajectory of the target is learned in the large-scale feature space, and then gradually focused on the small-scale feature extraction of the fine texture of the target. To improve the diversity of the generated samples, the system integrates a conditional variational autoencoder, and realizes the continuous adjustment of key parameters such as target size, motion speed, and radiation intensity by introducing random perturbation vectors in the latent space. In particular, the generation model also integrates the prior knowledge of the infrared physical imaging process, and ensures that the generated pseudo-targets conform to the Stefan-Boltzmann law through the constraint of the heat radiation equation. The finally output pseudo-sample images not only have realistic infrared characteristics, but also come with complete metadata tags, including information such as the equivalent blackbody temperature, motion parameters, and background complexity of the target, providing rich supervision signals for the training and evaluation of subsequent detection algorithms. Through the end-to-end training method, this generation system realizes the intelligent conversion from degraded samples to high-quality pseudo-samples, effectively solving the problem of difficult acquisition of real infrared small and weak target samples. Figure 4 The comparison schematic diagram between the generated pseudo-samples, the pseudo-samples generated by the CycleGAN method, and the pseudo-samples generated by the DCGAN method is shown. The method of this embodiment shows significant advantages in generating high-quality and highly consistent pseudo-datasets, especially 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, unmanned aerial vehicle defense systems, and various automated monitoring systems. Compared with the CycleGAN and DCGAN generation methods, the method of the present invention performs better in terms of the quality, stability, and consistency of the generated data. The images generated by the present invention are closer to the original images in terms of details and gray-scale distribution, and maintain a high degree of accuracy in terms of target scale and spatial structure. While CycleGAN and DCGAN have obvious deficiencies in generating complex scenes, especially in terms of gray-scale distribution, color offset, and spatial structure.

[0101] S5. Pseudo-dataset Comparative Detection Experiment: To verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model, the generated small target sample data was detected using the yolov10s, TLLCM, CNN, Faster R-CNN, and GAN algorithm models. The experiment was carried out on a high-performance computing platform equipped with NVIDIA Quadro RTX 4090 (32G video memory), using the Ubuntu 18.04 operating system environment and 8G of memory to ensure the stable operation of the system. The experiment constructed a complete GPU-accelerated computing environment and performed model training and verification based on the CUDA 12.0 parallel computing architecture and the PyTorch deep learning framework. In terms of the software environment, the flexibility and scalability of the PyTorch framework created favorable conditions for the rapid iteration and experimental verification of the model, while the optimized computing power of CUDA 12.0 significantly improved the model training efficiency, ensuring that the complete verification process of the algorithm could be completed within a reasonable time. Figure 5 For qualitative results.

[0102] S6. Pseudo-dataset Sample Control Experiment: To further verify the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model, a control experiment was conducted using the generated dataset and two classic infrared small and weak target pseudo-sample datasets. The experiment used Dataset Four and Dataset Five to train on the network architecture (YOLOv10s), and a unified evaluation metric was used to compare the performance of the model on the real infrared dataset. The evaluation metrics used IoU and false alarm rate as pixel-level evaluation metrics, and detection probability was used to evaluate the target-level performance. Furthermore, a control experiment was conducted using the generated pseudo-dataset and the IRSTD-1k Dataset Five, NUDT-SIRST Dataset Seven, NUST-SIRST Dataset Eight, and IRDST-simulation Dataset Nine. Through quantitative analysis, the effectiveness of the "low, slow, and small" moving target sample generation method based on the degradation model was further verified. Figure 6 For quantitative results.

[0103] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for generating samples of low, slow, and small moving targets based on a degradation model, characterized in that It includes the following steps: S1. Prepare the dataset: Use nine datasets. Among them, Dataset One, Dataset Two, Dataset Three, and Dataset Six are used for pseudo-sample generation; Dataset Four and Dataset Five are used for training the model in the comparative experiment; Dataset Five, Dataset Seven, Dataset Eight, and Dataset Nine are used for the comparative experiment with the generated pseudo-sample dataset; 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: It includes a target degradation process and a background degradation process. Input the target pseudo-samples and background pseudo-samples generated in Step S2 into this model to obtain a degraded sample library; S4. Construct a target pseudo-sample generation model: Input the degraded sample library in Step S3 into this model to generate infrared small and weak target pseudo-sample images; S5. Comparative detection experiment on the pseudo-dataset: Use multiple algorithm models to detect the generated pseudo-sample data; S6. Control experiment on the pseudo-dataset samples: Conduct a control experiment by comparing the generated dataset with four classic infrared small and weak target pseudo-sample datasets to verify the effectiveness of the generation method.

2. The method for generating samples of low, slow, and small moving targets based on a degradation model according to claim 1, wherein In the above S1: Dataset One is the MDvsFA dataset, Dataset Two is the SIRST-Aug dataset, and Dataset Three is the SIRST dataset; Dataset Four is the NUAA-SIRST dataset, and Dataset Five is the IRSD-1k dataset; Dataset Six is a self-made infrared dataset containing small and weak targets; Dataset Seven is the NUDT-SIRST dataset, Dataset Eight is the NUST-SIRST dataset, and Dataset Nine is the IRDST-simulation dataset.

3. The method for generating samples of low, slow, and small moving targets based on a degradation model according to claim 1, wherein The generative adversarial network with controllable azimuth angle in the above S2 includes: Generator: Used to extract and fuse target features from input images at adjacent azimuth angles to generate pseudo-samples between the two azimuth angles; Similarity discriminator: Used to evaluate the similarity between the generated image and the real image; Azimuth predictor: Used to predict the azimuth angle of the generated image; Loss function: The generator uses an adversarial loss function, a content loss function, a perceptual loss function, a synthesis loss function, and a total variation loss function; The similarity discriminator uses an adversarial loss function.

4. The method for generating low, slow, and small moving target samples based on a degradation model according to claim 1, characterized in that, The image degradation model in the above S3 includes: Target degradation: Degrade using an isotropic / anisotropic Gaussian blur kernel and a multi-directional motion blur function; Background degradation: Degrade using a noise degradation module, including Gaussian noise, Poisson noise, JPEG compression noise, and sensor noise.

5. The method for generating samples of low, slow, and small moving targets based on a degradation model according to claim 1, wherein The pseudo-sample generation model in the above S4 adopts: Quantity constraint: Control the number of targets added to the background image within a preset range; Position constraint: Ensure that the position layout of the targets in the background image is reasonable.

6. The method for generating samples of low, slow, and small moving targets based on a degradation model according to claim 1, wherein The detection algorithm models adopted in the above S5 include: YOLOv10s, TLLCM, CNN, Faster R-CNN, and GAN, where the model training is based on Dataset Three and Dataset Four.

7. The method for generating low, slow, and small moving target samples based on a degradation model according to claim 1, wherein The control experiment in the above S6 includes: Compare the generated pseudo-sample dataset with the IRSTD-1k dataset V, NUDT-SIRST dataset VII, NUST-SIRST dataset VIII, and IRDST-simulation dataset IX; Use the YOLOv10s model trained based on dataset IV and dataset V for detection and verification.

Images