Infrared small target detection method and system based on visual autoregression teacher model

Abstract

The invention discloses an infrared small target detection method and system based on a visual autoregression teacher model, and relates to the technical field of computer vision and image processing. The method mainly comprises the following steps: constructing and training a VAR teacher model, extracting multi-scale features by using a frozen pre-trained visual autoregression model, and adapting general features to an infrared small target detection task through a lightweight adapter module; using the trained teacher model to generate a high-quality pseudo label for the target domain unlabeled infrared image, and carrying out confidence threshold screening and non-maximum suppression; training a lightweight student detector by using a joint loss function based on a self-training framework in combination with the source domain labeling data and the target domain pseudo-tag; and updating the average teacher model through an index moving average mechanism for final reasoning and evaluation. According to the method, the problems of low precision, low robustness and weak generalization ability in cross-domain infrared small target detection are solved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, specifically to an infrared small target detection method and system based on a visual autoregressive teacher model. Background Technology

[0002] Infrared small target detection has wide applications in emergency early warning, security monitoring, and drone navigation. However, this task faces significant challenges: First, targets in infrared images are typically small in size, have weak texture features, and low signal-to-noise ratios; second, labeling large-scale, high-quality infrared image data requires specialized knowledge and incurs substantial costs, leading to a scarcity of labeled data; third, due to differences in imaging equipment, scenes, and environmental conditions, there are significant distributional differences between different domains, causing a sharp performance drop when a model trained on one dataset is directly applied to another dataset—a problem known as domain adaptation.

[0003] Existing solutions mainly include methods based on traditional filtering and methods based on deep learning. Traditional methods (such as Top-Hat and LCM) have limited generalization ability in complex contexts. While deep learning methods offer excellent performance, they heavily rely on large amounts of labeled data. Although some studies have attempted to use Generative Adversarial Networks (GANs) for domain transformation or utilize unsupervised methods, these methods often suffer from unstable training or difficulty in maintaining the ability to capture the fine structure of small targets.

[0004] Therefore, there is an urgent need for a cross-domain infrared small target detection method that can effectively utilize a small amount of source domain labeled data and a large amount of target domain unlabeled data to achieve high precision, high robustness, and strong generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide an infrared small target detection method and system based on a visual autoregressive teacher model, so as to solve the problems of low accuracy, low robustness and weak generalization ability of infrared small target detection under cross-domain conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An infrared small target detection method based on a visual autoregressive teacher model includes the following steps:

[0008] S1: Construct a VAR teacher model and train it using source domain labeled data;

[0009] S2: Use the trained VAR teacher model to generate pseudo-labels for unlabeled data in the target domain;

[0010] S3: Based on a self-training framework, a student detector is trained by combining source domain supervised data and target domain pseudo-label data.

[0011] S4: Update the average teacher model using an exponential moving average mechanism and evaluate its performance using evaluation metrics.

[0012] Furthermore, S1 includes:

[0013] S11: Perform standardization preprocessing on the input infrared image to obtain the normalized image I. norm ;

[0014] S12: Input the preprocessed image into the frozen pre-trained visual autoregressive model for multi-scale feature extraction, and obtain the multi-scale feature representation F. var ;

[0015] S13: The multi-scale features are processed and transformed through the adapter module, and the adapted feature pyramid F is output. adapt ;

[0016] S14: Using the detection head based on F adapt Output the prediction results and train the adapter module and detection head on the source domain labeled data.

[0017] Furthermore, the adapter module includes a channel adjustment layer, a multi-scale spatial attention mechanism, and an improved feature pyramid network;

[0018] The calculation formula for the multi-scale spatial attention mechanism is as follows:

[0019]

[0020] Where σ is the Sigmoid activation function, used to generate attention weights between 0 and 1; ⊙ represents the element-wise multiplication operation, using the increment operation to avoid complete suppression of important features; Conv 7×7 This indicates that the 7×7 convolution adjusts the channel dimensions.

[0021] Furthermore, S2 includes:

[0022] S21: Apply a weak enhancement transform to the target domain image;

[0023] S22: Use the VAR teacher model to predict weakly enhanced images;

[0024] S23: The prediction results are filtered using a dynamically adjusted confidence threshold and non-maximum suppression (NMS) to generate pseudo-labels (PseudoLabels).

[0025] Furthermore, S23 specifically refers to:

[0026] A preliminary screening is performed using a dynamically adjusted confidence threshold, expressed as follows:

[0027]

[0028] Where, σ init The initial confidence threshold is typically set to 0.8 to ensure high quality of initial pseudo-labels. The QualityFactor is a quality adjustment factor that is dynamically adjusted based on the overall quality of the current pseudo-labels. T total Total training steps;

[0029] Then, an improved nonmaximum suppression algorithm is applied to remove duplicate detections and low-quality predictions, generating pseudo-labels (PseudoLabels). j The expression is:

[0030] PseudoLabels j =(bbox) k ,cls k ,conf k )|conf k >σ t NMS (bbox) k IoU threshold =0.5)

[0031] Among them, bbox k For the k-th bounding box, cls k For category labels, conf k The confidence score is 0.5. NMS is a non-maximum suppression operation, and the IoU threshold for NMS is set to 0.5.

[0032] Furthermore, S3 includes:

[0033] S31: Construct a lightweight student detector;

[0034] S32: Apply a strong enhancement transform to the target domain image, the expression is:

[0035] T strong (I)=Compose[RandomRotate(),ColorJitter(brightness=0.4)]

[0036]

[0037] Where I represents the original input image; ColorJitter is a common data augmentation technique that creates new data by randomly fine-tuning the color attributes of an image; brightness = 0.4 is a key parameter, representing that the brightness of the image will randomly vary within a large range;

[0038] S33: The student model is trained using a multi-task joint loss function, expressed as follows:

[0039] L total =L sup +λ unsup ×L unsup

[0040] Among them, L sup For source domain supervision loss, L unsup The loss is the supervision loss for pseudo-labels in the target domain.

[0041] Furthermore, the average teacher model updated using the exponential moving average mechanism described in S4 is expressed as follows:

[0042]

[0043] in, Let be the average teacher parameter at step t. Let be the student model parameters at step t, and α be the EMA decay rate, ranging from 0.99 to 0.999.

[0044] Furthermore, the evaluation metrics in S4 include mean precision (mAP) and recall.

[0045] Recall, Precision, and F1 score are calculated using the following formulas:

[0046]

[0047] Where C is the number of categories, AP c For the average precision of class c, P c (R c The curve represents the precision-recall ratio, where TP represents true positives and FN represents false negatives.

[0048] mAP at different IoU thresholds is as follows:

[0049]

[0050] Another objective of this invention is to provide an infrared small target detection system based on a visual autoregressive teacher model, wherein the system executes the aforementioned infrared small target detection method based on a visual autoregressive teacher model, comprising:

[0051] VAR Teacher Model Construction and Training Unit: Used to construct a teacher model containing a frozen pre-trained visual autoregressive model and a trainable adapter module, and to train it using source domain labeled data. It includes a model loading module, an adapter module, and a training module.

[0052] Pseudo-label generation unit: Utilizes a trained teacher model to generate high-quality pseudo-labels for unlabeled data in the target domain;

[0053] Student Model Training and Average Teacher Update Unit: Based on source domain data and pseudo-labels, a student detector is trained through a joint loss function, and an average teacher model is updated using an exponential moving average mechanism for inference.

[0054] Evaluation Unit: The final model's performance is evaluated using metrics such as mean precision (mAP) and recall.

[0055] Furthermore, the model loading module loads the pre-trained visual autoregressive model and freezes its parameters;

[0056] The adapter module consists of a channel adjustment layer and a feature pyramid network (FPN), which is responsible for converting the multi-scale general features extracted by the VAR teacher model into a feature pyramid suitable for infrared small target detection.

[0057] The training module trains the parameters of the adapter module only on the source domain labeled data, and completes the adaptation training of the VAR teacher model by minimizing the classification loss and regression loss.

[0058] In summary, the present invention has at least one of the following beneficial technical effects:

[0059] By utilizing models pre-trained on large-scale datasets, it provides rich and general visual feature priors, laying a solid foundation for small object detection and possessing powerful feature representation capabilities;

[0060] The framework of generating pseudo-labels through a teacher model and self-training through a student model effectively utilizes unlabeled data in the target domain, solves the problem of cross-domain differences, and demonstrates strong generalization ability of the model.

[0061] By employing multiple strategies such as weak enhancement, dynamic threshold filtering, and quality verification, the reliability of pseudo-labels is ensured, the accumulation of erroneous labels during training is avoided, and the system exhibits strong robustness.

[0062] The student model adopts a lightweight design to ensure reasoning speed and high efficiency;

[0063] The average teacher model updated through the EMA mechanism has high robustness and high accuracy. Attached Figure Description

[0064] Figure 1 This is a flowchart of the present invention;

[0065] Figure 2 This is a diagram of the VAR teacher model architecture of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0067] This embodiment provides an infrared small target detection method based on a visual autoregressive teacher model, such as... Figure 1 As shown, it includes the following steps:

[0068] S1: Construct a VAR teacher model and train it using source domain labeled data;

[0069] S2: Use the trained VAR teacher model to generate pseudo-labels for unlabeled data in the target domain;

[0070] S3: Based on a self-training framework, a student detector is trained by combining source domain supervised data and target domain pseudo-label data.

[0071] S4: Update the average teacher model using an exponential moving average mechanism and evaluate its performance using evaluation metrics.

[0072] The following description is based on specific embodiments.

[0073] S1: Construct a VAR teacher model and train it using source domain labeled data.

[0074] The VAR teacher model includes: a frozen pre-trained visual autoregressive model, an adapter module, and a detection head. The visual autoregressive model, pre-trained on large-scale visual data, is loaded.

[0075] The model was thoroughly pre-trained on multiple large datasets such as ImageNet and COCO, possessing powerful general visual representation learning capabilities. All parameters of the visual autoregressive model were frozen and did not participate in gradient updates or parameter optimization throughout the training process, thus fully preserving the rich knowledge and powerful capabilities learned by the pre-trained model. Specifically, it consists of the following four steps:

[0076] 1. Perform standardization preprocessing on the input infrared image to obtain the normalized image I. norm .

[0077] First, an infrared small target detection dataset is constructed, which includes diverse infrared image data collected from different scenes, times, and devices.

[0078] Then, comprehensive preprocessing operations are performed on the input infrared images, including data format unification, numerical range normalization, noise filtering, and contrast enhancement, to ensure the quality and consistency of the input data.

[0079] The input infrared image undergoes a standardization preprocessing operation, expressed as:

[0080]

[0081] Among them, I norm The normalized input image ensures that the pixel values ​​are within a reasonable range, μ I and σ I , respectively, are the statistical mean and standard deviation of the infrared image dataset, obtained through statistical analysis of a large number of infrared images, and ∈ is a protection constant to prevent division by zero, usually taken as 1e-8.

[0082] 2. Input the preprocessed image into the frozen pre-trained visual autoregressive model for multi-scale feature extraction to obtain the multi-scale feature representation F. var .

[0083] Multi-scale feature extraction in the VAR teacher model is the foundation and key to the entire detection process. The preprocessed image is input into the frozen visual autoregressive model for multi-scale feature extraction, expressed as:

[0084]

[0085] The model can extract feature representations at different scales and semantic levels from its multiple intermediate layers, from low-level edge texture features to high-level semantic abstract features, providing a rich information foundation for subsequent detection tasks.

[0086] 3. The multi-scale features are processed and transformed using an adapter module to output the adapted feature pyramid F. adapt ;

[0087] The meticulously designed adapter module serves as a crucial bridge connecting general VAR features and infrared detection tasks. The adapter comprises several key components, including a channel adjustment layer, a multi-scale spatial attention mechanism, and an improved feature pyramid network. Specifically, the channel adjustment layer handles dimensional alignment, the multi-scale spatial attention mechanism enhances important regions, and the improved feature pyramid network handles multi-scale feature fusion. The number of parameters in the entire adapter is kept within a reasonable range, achieving effective feature transformation without introducing excessive computational overhead.

[0088] First, the number of channels in the VAR features is adjusted to a dimension suitable for the detection task using a channel adjustment layer, expressed as:

[0089]

[0090] in, Features extracted from the i-th layer of VAR For the adjusted features, This corresponds to a 1×1 convolutional layer.

[0091] The ResidualConnection design is adopted here, which can effectively prevent the loss of feature information during the transformation process and accelerate the training and convergence of the network.

[0092] Then, a multi-scale spatial attention mechanism is applied to automatically identify and enhance the regions and features most important for small object detection in the image. The calculation formula is as follows:

[0093]

[0094] Where σ is the Sigmoid activation function, used to generate attention weights between 0 and 1; ⊙ represents the element-wise multiplication operation, using the increment operation to avoid complete suppression of important features; Conv 7×7 This indicates that the 7×7 convolution adjusts the channel dimensions.

[0095] Furthermore, by fusing multi-scale features through an improved feature pyramid network, an effective combination of semantic information and spatial details is achieved, expressed as:

[0096]

[0097] Here, FPN stands for Feature Pyramid Network.

[0098] 4. Using the detection head based on F adapt Output the prediction results and train the adapter module and detection head on the source domain labeled data.

[0099] The key steps are to use the detection head to make predictions based on features output by the adapter, and to train the adapter module and the detection head only on the source domain labeled data. The source domain training employs a carefully designed multi-task loss function that can simultaneously optimize classification accuracy, localization accuracy, and detection confidence.

[0100] The multi-task loss function is expressed as follows:

[0101] L source =L cls +λ reg ×L reg +λ focal ×L focal

[0102] Among them, the classification loss L cls To improve generalization ability, a label-smoothed cross-entropy loss is used, expressed as:

[0103]

[0104] Regression loss L reg Improving positioning accuracy by using IoU-aware SmoothL1 loss, the expression is:

[0105]

[0106] Focal loss L focal To handle the problem of imbalanced positive and negative samples, the expression is:

[0107]

[0108] Where, λ reg and λ focal The loss weight coefficients are determined through grid search and validation set performance, typically set to 1.0 and 2.0 respectively; α i γ is the balance factor, and γ is the focusing parameter, which is usually set to 2.0.

[0109] like Figure 2 As shown, the VAR teacher model architecture constructed in this embodiment demonstrates the precise cooperation and coordination between its various components. The model comprises four main parts: an input preprocessing module, a frozen VAR backbone network, a trainable adapter module, and a detection head. The input preprocessing module is responsible for image standardization and enhancement; the VAR backbone network extracts rich multi-scale feature representations from multiple intermediate layers; the adapter module contains six channel adjustment layers at different scales, each equipped with a dedicated attention mechanism and residual connections; the feature pyramid network employs a top-down and bottom-up bidirectional fusion strategy, which can fully utilize the complementary information of features at different levels; the detection head contains three branches: classification, regression, and confidence, responsible for target category prediction, bounding box localization, and detection quality evaluation, respectively. The entire system is implemented through the following mathematical process:

[0110] Input image I → VAR feature extraction F var →Adapter processing F adapt →Detector head outputs Y pred .

[0111] S2: Generate pseudo-labels for the unlabeled data in the target domain using the trained VAR teacher model, which consists of the following 3 steps:

[0112] 1. Apply a weak enhancement transform to the target domain image;

[0113] First, prepare an unlabeled dataset for the target domain. These images typically have characteristics different from the source domain data, including different sensor types, imaging conditions, background environments, and target distributions. Nt This indicates the number of samples in the target domain dataset, which is usually much larger than the number of labeled data in the source domain.

[0114] Secondly, in order to maintain the accuracy and reliability of the pseudo-labels, a carefully designed weak enhancement transformation is applied to each target domain image. The intensity and type of the weak enhancement are carefully adjusted to ensure that the true location, size and appearance features of the target are not significantly changed.

[0115] The design of the weak enhancement transform fully considers the special properties of infrared images and the sensitivity requirements for small target detection. Its expression is:

[0116] T weak (I)=ColorJitter(I,brightness=0.1)

[0117]

[0118] Where I is the original input image; ColorJitter is a common data augmentation technique that creates new data by randomly fine-tuning the color attributes of an image; brightness=0.1 is a key parameter that limits the image brightness to be fine-tuned only randomly, and the jitter amplitude is very small.

[0119] Weak enhancement includes only a 30% probability of horizontal flipping and very slight brightness and contrast adjustments, with the variation controlled within 5%. This increases the slight diversity of the data without affecting the true features of the target and the accuracy of the annotation.

[0120] 2. Use the VAR teacher model to predict weakly enhanced images.

[0121] The VAR teacher model trained on the source domain is used to perform prediction inference on the weakly enhanced target domain image to obtain preliminary detection results including bounding boxes, categories, and confidence levels.

[0122] 3. The prediction results are filtered using a dynamically adjusted confidence threshold and non-maximum suppression (NMS) to generate pseudo-labels (PseudoLabels).

[0123] To ensure the high quality of pseudo-labels, a multi-stage screening and verification strategy is adopted, including confidence threshold screening, temporal consistency verification, spatial rationality check, and category balance adjustment.

[0124] After the VAR teacher model predicts weakly enhanced images, it first performs preliminary screening using a dynamically adjusted confidence threshold, the expression of which is:

[0125]

[0126] Where, σ init The initial confidence threshold is typically set to 0.8 to ensure high quality of initial pseudo-labels. The QualityFactor is a quality adjustment factor that is dynamically adjusted based on the overall quality of the current pseudo-labels. T total Total training steps;

[0127] Then, an improved nonmaximum suppression algorithm is applied to remove duplicate detections and low-quality predictions, generating pseudo-labels (PseudoLabels). j The expression is:

[0128] PseudoLabels j =(bbox) k ,cls k ,conf k )|conf k >σ t NMS (bbox) k IoU threshold =0.5)

[0129] Among them, bbox k For the k-th bounding box, cls k For category labels, conf k The confidence score is used for Non-Maximum Suppression (NMS). The IoU threshold for NMS is set to 0.5, which effectively removes duplicate detections without over-suppressing adjacent real targets.

[0130] Quality checks include multiple dimensions such as bounding box rationality verification, category consistency check, and spatial distribution rationality assessment.

[0131] S3: Based on a self-training framework, the student detector is trained by combining source domain supervised data and target domain pseudo-label data, which consists of the following three steps:

[0132] 1. Construct a lightweight student detector.

[0133] Building an efficient, lightweight student detector typically employs MobileNet, EfficientNet, or a custom lightweight network as its backbone architecture, maximizing inference efficiency and deployment flexibility while ensuring detection performance.

[0134] 2. To improve the robustness and generalization ability of the model, stronger data augmentation transformations are applied to the target domain image, including various techniques such as geometric transformation, color transformation, noise addition, and occlusion simulation.

[0135] Applying a strong enhancement transform to the target domain image, the expression is:

[0136] T strong(I)=Compose[RandomRotate(),ColorJitter(brightness=0.4)]

[0137]

[0138] Where I represents the original input image; ColorJitter is a common data augmentation technique that creates new data by randomly fine-tuning the color attributes of an image; brightness = 0.4 is a key parameter, representing that the brightness of the image will vary randomly within a large range.

[0139] 3. The student model is trained using a carefully designed multi-task joint loss function, which can effectively balance multiple learning objectives such as source domain supervised learning, target domain self-supervised learning, knowledge distillation, and consistency constraints.

[0140] The student model is trained using a multi-task joint loss function, expressed as follows:

[0141] L total =L sup +λ unsup ×L unsup

[0142] Source domain supervision loss L sup expression:

[0143]

[0144] Target domain pseudo-label supervision loss L unsup expression:

[0145]

[0146] Where, N s and N t y and b represent the number of samples in the source and target domains, respectively, and the category and bounding box label, respectively.

[0147] A progressive weight adjustment strategy is adopted during training. Initially, it mainly relies on source domain supervision signals, and gradually increases the weight of pseudo-labels as training progresses, ultimately achieving effective domain adaptation.

[0148] S4: Update the average teacher model using an exponential moving average mechanism and evaluate its performance using evaluation metrics.

[0149] First, the average teacher model parameters are updated using an exponential moving average mechanism.

[0150] An average teacher model with the exact same architecture as the student model is created and used as the primary model for final inference and evaluation. The parameters of the average teacher model are updated using a dynamically adaptive exponential moving average strategy after each update of the student model parameters.

[0151] The average teacher model is used for final inference and assessment, and is updated using the EMA (Exponential Moving Average) mechanism, expressed as:

[0152]

[0153] in, Let be the average teacher parameter at step t. Let be the student model parameters at step t, and α be the EMA decay rate, ranging from 0.99 to 0.999.

[0154] The parameter updates of the average teacher model do not rely on gradient calculation, but are obtained through a weighted average of historical parameters. This approach can effectively smooth parameter fluctuations during training and provide more stable and robust model performance. At the same time, a performance monitoring and rollback mechanism is introduced so that when the model performance drops abnormally, it can automatically recover to the historical best state.

[0155] Secondly, performance evaluation is conducted using evaluation metrics.

[0156] A comprehensive performance evaluation was performed on specially constructed validation and test sets using the trained average teacher model. Multi-dimensional, multi-level evaluation metrics were calculated, employing a progressive training strategy that initially reduces the weight of pseudo-labels and gradually increases it as training progresses. Evaluation metrics included mean AP, recall, precision, and F1 score, calculated using the following formulas:

[0157]

[0158] Where C is the number of categories, AP c For the average precision of class c, P c (R c The curve represents the precision-recall ratio, where TP represents true positives and FN represents false negatives.

[0159] mAP at different IoU thresholds is as follows:

[0160]

[0161] In this process, detailed error analysis and case studies should be conducted to identify the strengths and weaknesses of the model and provide guidance for further improvement; the effectiveness and advancement of the method of this invention should be fully verified by comparing multiple baseline methods and the latest technologies; and the average teacher model should be used to perform inference on the validation set and the evaluation index should be calculated.

[0162] Table 1 Comparison of Infrared Small Target Detection Performance

[0163] Methods mAP@0.5 Recall FPS Parameters (M) YOLOv7-tiny 63.2 76.5 145 6.0 DNANet 86.1 98.7 48 4.7 ALCNet 75.7 96.5 52 0.9 Ours 87.5 99.1 50 4.5

[0164] The experimental results of this invention on the SIRST infrared small target detection dataset are shown in Table 1. The data in Table 1 clearly demonstrate the significant advantages of the method of this invention (Ours) in terms of accuracy and detection capability compared to other state-of-the-art methods, while maintaining good efficiency. Specifically:

[0165] 1. High accuracy: The present invention achieves 87.5% accuracy in the core detection accuracy metric mAP@0.5 (mean accuracy, IoU threshold set to 0.5), significantly higher than other comparative methods (YOLOv7-tiny: 63.2%, DNANet: 86.1%, ALCNet: 75.7%). This indicates that the detection bounding box generated by the present invention has a high degree of consistency with the actual target position, resulting in good overall detection accuracy.

[0166] 2. Excellent Recall: This invention achieves a recall rate of 99.1%, slightly higher than the second-place DNANet (98.7%). This means that this invention can detect almost all real small targets in an image, with a low false negative rate. This is crucial for applications with high requirements such as security and early warning systems.

[0167] 3. Good balance between efficiency and accuracy: In terms of inference speed, the FPS (frames per second) of this invention is 50, which is lower than that of the lightweight YOLOv7-tiny (145), but on par with DNANet (48) and ALCNet (52), which are also high-performance models, and can meet the needs of real-time detection; In terms of model complexity, the number of parameters of this invention is 4.5M (millions), which is comparable to DNANet (4.7M) and far lower than that of YOLOv7-tiny (6.0M). Moreover, the higher number of parameters than ALCNet (0.9M) has resulted in a huge improvement in its mAP and Recall, which shows that its model design is efficient and effective.

[0168] Therefore, the data in Table 1 proves that the infrared small target detection method based on a visual autoregressive teacher model described in this invention successfully achieves its inventive purpose and effectively solves the problems of "low accuracy, low robustness, and weak generalization ability". Without sacrificing too much inference speed, it simultaneously achieves high detection accuracy (mAP) and excellent target detection capability (Recall), surpassing the performance of other listed advanced models and demonstrating strong comprehensive advantages.

[0169] This embodiment also provides an infrared small target detection system based on a visual autoregressive teacher model. This system executes the infrared small target detection method based on a visual autoregressive teacher model described in Embodiment 1, including a VAR teacher model construction and training unit, a pseudo-label generation unit, a student model training and average teacher update unit, and an evaluation unit. The specific details are as follows:

[0170] The VAR teacher model construction and training unit is used to build a teacher model that includes a frozen pre-trained visual autoregressive model and a trainable adapter module, and train it using source domain labeled data. It includes a model loading module, an adapter module, and a training module.

[0171] The model loading module loads the pre-trained visual autoregressive model and freezes its parameters; the adapter module consists of a channel adjustment layer and a feature pyramid network (FPN), which is responsible for converting the multi-scale general features extracted by the VAR teacher model into a feature pyramid suitable for infrared small target detection; the training module trains the parameters of the adapter module only on the source domain labeled data, and completes the adaptation training of the VAR teacher model by minimizing the classification loss and regression loss.

[0172] The pseudo-label generation unit uses a trained teacher model to generate high-quality pseudo-labels for unlabeled data in the target domain.

[0173] The student model training and average teacher update unit trains a student detector based on source domain data and pseudo-labels using a joint loss function, and uses an exponential moving average mechanism to update an average teacher model that is ultimately used for inference.

[0174] The evaluation unit uses metrics such as mean precision (mAP) and recall to evaluate the performance of the final model.

[0175] The core of this invention lies in the teacher-student self-training framework. First, a teacher model is constructed using a VAR teacher model building unit. This model feeds the input infrared image into a frozen VAR backbone network to extract multi-scale features F. var Next, the adapter module adjusts the feature channels through 1x1 convolution and fuses them using FPN, outputting a feature pyramid F relevant to the detection task. adapt The detection head outputs prediction results based on this feature pyramid. Then, the pseudo-label generation unit uses the trained teacher model to predict the target domain data and generates high-quality pseudo-labels through confidence thresholding and non-maximum suppression (NMS). Finally, the student model training unit uses the joint loss function L... total Training is performed, incorporating the supervised loss L from the source domain. sup The pseudo-label supervision loss L in the target domain unsupThe knowledge is distilled from a robust teacher model into a lightweight student model. Simultaneously, the average teacher update unit continuously updates a more stable average teacher model using exponential moving average (EMA) for final evaluation and deployment. Through this process, high-precision, highly robust, and generalizable infrared small target detection is achieved.

[0176] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0177] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0178] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0179] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0180] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A method for detecting small infrared targets based on a visual autoregressive teacher model, characterized in that, Includes the following steps: S1: Construct a VAR teacher model and train it using source domain labeled data; S2: Use the trained VAR teacher model to generate pseudo-labels for unlabeled data in the target domain; S3: Based on a self-training framework, a student detector is trained by combining source domain supervised data and target domain pseudo-label data. S4: Update the average teacher model using an exponential moving average mechanism and evaluate its performance using evaluation metrics.

2. The infrared small target detection method based on a visual autoregressive teacher model according to claim 1, characterized in that, S1 includes: S11: Perform standardization preprocessing on the input infrared image to obtain the normalized image I. norm ; S12: Input the preprocessed image into the frozen pre-trained visual autoregressive model for multi-scale feature extraction, and obtain the multi-scale feature representation F. var ; S13: The multi-scale features are processed and transformed through the adapter module to output the adapted feature pyramid F. adapt ; S14: Using the detection head based on F adapt Output the prediction results and train the adapter module and detection head on the source domain labeled data.

3. The infrared small target detection method based on a visual autoregressive teacher model according to claim 2, characterized in that: The adapter module includes a channel adjustment layer, a multi-scale spatial attention mechanism, and an improved feature pyramid network. The calculation formula for the multi-scale spatial attention mechanism is as follows: Where σ is the Sigmoid activation function, used to generate attention weights between 0 and 1; ⊙ represents the element-wise multiplication operation, using the increment operation to avoid complete suppression of important features; Conv 7×7 This indicates that the 7×7 convolution adjusts the channel dimensions.

4. The infrared small target detection method based on a visual autoregressive teacher model according to claim 1, characterized in that, S2 include: S21: Apply a weak enhancement transform to the target domain image; S22: Use the VAR teacher model to predict weakly enhanced images; S23: The prediction results are filtered using a dynamically adjusted confidence threshold and non-maximum suppression (NMS) to generate pseudo-labels (PseudoLabels).

5. The infrared small target detection method based on a visual autoregressive teacher model according to claim 4, characterized in that, S23 specifically refers to: A preliminary screening is performed using a dynamically adjusted confidence threshold, expressed as follows: Where, σ init The initial confidence threshold is typically set to 0.8 to ensure high quality of initial pseudo-labels. The QualityFactor is a quality adjustment factor that is dynamically adjusted based on the overall quality of the current pseudo-labels. T total Total training steps; Then, an improved nonmaximum suppression algorithm is applied to remove duplicate detections and low-quality predictions, generating pseudo-labels (PseudoLabels). j The expression is: PseudoLabels j =(bbox k ,cls k ,conf k )|conf k >σ t ,NMS(bbox k ,IoU threshold =0.5) Among them, bbox k For the k-th bounding box, cls k For category labels, conf k The confidence score is 0.

5. NMS is a non-maximum suppression operation, and the IoU threshold for NMS is set to 0.

5.

6. The infrared small target detection method based on a visual autoregressive teacher model according to claim 1, characterized in that, S3 include: S31: Construct a lightweight student detector; S32: Apply a strong enhancement transform to the target domain image, the expression is: T strong (I)=Compose[RandomRotate(),ColorJitter(brightness=0.4)] Where I represents the original input image; ColorJitter is a common data augmentation technique that creates new data by randomly fine-tuning the color attributes of an image; brightness = 0.4 is a key parameter, representing that the brightness of the image will randomly vary within a large range; S33: The student model is trained using a multi-task joint loss function, expressed as follows: L total =L sup +λ unsup ×L unsup Among them, L sup For source domain supervision loss, L unsup The loss is the supervision loss for pseudo-labels in the target domain.

7. The infrared small target detection method based on a visual autoregressive teacher model according to claim 1, characterized in that, The average teacher model updated using the exponential moving average mechanism described in S4 is expressed as follows: in, Let be the average teacher parameter at step t. Let be the student model parameters at step t, and α be the EMA decay rate, ranging from 0.99 to 0.

999.

8. The infrared small target detection method based on a visual autoregressive teacher model according to claim 1, characterized in that, The evaluation metrics in S4 include mean accuracy (mAP), recall, precision, and F1 score, calculated using the following formulas: Where C is the number of categories, AP c P represents the average precision of class c. c (R c The curve represents the precision-recall ratio, where TP represents true positives and FN represents false negatives. mAP at different IoU thresholds is as follows:

9. An infrared small target detection system based on a visual autoregressive teacher model, characterized in that, The system executes an infrared small target detection method based on a visual autoregressive teacher model as described in any one of claims 1 to 8, comprising: VAR Teacher Model Construction and Training Unit: Used to construct a teacher model containing a frozen pre-trained visual autoregressive model and a trainable adapter module, and to train it using source domain labeled data. It includes a model loading module, an adapter module, and a training module. Pseudo-label generation unit: Utilizes a trained teacher model to generate high-quality pseudo-labels for unlabeled data in the target domain; Student Model Training and Average Teacher Update Unit: Based on source domain data and pseudo-labels, a student detector is trained through a joint loss function, and an average teacher model is updated using an exponential moving average mechanism for inference. Evaluation Unit: The final model's performance is evaluated using metrics such as mean precision (mAP) and recall.

10. The infrared small target detection system based on a visual autoregressive teacher model according to claim 9, characterized in that: The model loading module loads the pre-trained visual autoregressive model and freezes its parameters; The adapter module consists of a channel adjustment layer and a feature pyramid network (FPN), which is responsible for converting the multi-scale general features extracted by the VAR teacher model into a feature pyramid suitable for infrared small target detection. The training module trains the parameters of the adapter module only on the source domain labeled data, and completes the adaptation training of the VAR teacher model by minimizing the classification loss and regression loss.

Images