An infrared dim small target detection method based on cross-domain feature migration

Abstract

This invention belongs to the field of target detection technology, specifically a method for detecting small infrared targets based on cross-domain feature transfer. A novel detection network, abbreviated as SA-YOLO, is proposed. First, inspired by the idea of ​​multispectral feature transfer, a position-guided enhancement module is designed. Through a triple structure of position-aware attention, multi-scale saliency fusion, and position-guided filtering, the infrared spectral feature response of small targets is effectively enhanced. Second, a spectral-spatial awareness module is proposed, employing a progressive hole rate and a dual-branch attention mechanism to achieve robust modeling of multi-scale spatial features. Then, an innovative sensitivity-stabilized loss function is proposed, mitigating the problem of a sharp drop in scale sensitivity for small targets through dynamic penalty weights and composite normalization. Finally, an infrared small target dataset called ISUAV-D is collected. Experimental results show that SA-YOLO exhibits significant advantages in environments with strong light interference and in small target detection scenarios.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically to a method for detecting weak infrared targets based on cross-domain feature transfer. Background Technology

[0002] In complex environments such as military reconnaissance and security monitoring, infrared imaging systems can effectively overcome the limitations of visible light sensors, achieving stable detection of distant, low signal-to-noise ratio targets. However, small infrared targets are usually composed of only a few pixels, lacking obvious shape and texture features, and are easily affected by complex backgrounds such as clouds and strong light. Existing methods are difficult to achieve high-precision detection of large targets that surpass those of visible light. Therefore, finding a method that can effectively transfer visible light-trained pre-networks to infrared detection tasks is an urgent problem to be solved. This invention mainly studies how to overcome the domain shift problem between visible light and infrared bands by using modular design of infrared spectral characteristics, while retaining the powerful feature extraction capabilities of the pre-trained network.

[0003] To address the aforementioned issues, traditional infrared weak target detection techniques are primarily based on manually designed features and prior knowledge. While these techniques perform well against uniform backgrounds, they rely on complex matrix operations, resulting in significant computational overhead when processing high-resolution images. Thanks to deep learning and massive data-driven approaches, detection based on convolutional neural networks has achieved numerous breakthroughs. Mainstream techniques include single-stage detectors (YOLO series) and two-stage detectors. Although existing single-stage and two-stage detectors have improved the network's feature extraction capabilities by introducing attention mechanisms and other methods, they have not studied the issue of infrared cross-domain feature transfer. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides an infrared weak target detection method based on cross-domain feature transfer. This method solves the problem of low detection accuracy caused by directly using visible light detection models due to challenges such as low signal-to-noise ratio and strong background interference for infrared weak targets. It also solves the problem of a sharp drop in the scale sensitivity of infrared weak targets when training regression algorithms using the traditional IoU loss function.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0008] An infrared weak target detection method based on cross-domain feature transfer specifically includes the following steps:

[0009] S1. Prepare two public datasets: Dataset 1 and Dataset 2, both of which contain training, validation and test sets.

[0010] S2, Prepare a self-built dataset: including data acquisition, data preprocessing, and data partitioning; Data 3 includes a training set, a validation set, and a test set;

[0011] S3, Design the Position Guidance Enhancement Module PSE_Former: PSE_Former includes the Position Awareness Module PAM, the Multi-Scale Saliency Fusion Module MSS, and the Position Guidance Filtering Module PGG; PAM, MSS, and PGG are cascaded in PSE_Former;

[0012] S4, Design Spectrum-Spatial Awareness Module SSPM: SSPM includes a progressive void rate grouping strategy and a local-global dual-path attention module LGAM;

[0013] S5, Design the sensitivity-stable loss function DS-IoU: DS-IoU achieves robust optimization of target scale differences through dynamic penalty weights and a composite normalization mechanism;

[0014] S6, Network Reconstruction and Model Training: Reconstruct a new network SA-YOLO on the basic YOLO network using PSE_Former from step S3 and SSPM from step S4; Train the model using the training sets of the three datasets from steps S1 and S2, and the DS-IoU loss and classification loss from step S5, while fine-tuning the model using the validation sets of the three datasets to further improve the cross-domain feature transfer detection performance from visible light to infrared.

[0015] S7, Model Testing and Evaluation: After the fine-tuning in step S6 is completed, solidify the fine-tuned network parameters and save the final weak target detection model; use the test set in steps S1 and S2 to test the model effect from a qualitative and quantitative perspective. Real infrared images can be directly input into the trained end-to-end model to obtain result images labeled with detection boxes and location information.

[0016] Furthermore, the specific acquisition methods and data structures of the NUST-SIRST dataset and the SIRST-5k dataset in S1 can be found in the existing descriptions.

[0017] Furthermore, in step S2, the data acquisition, data preprocessing, and data partitioning methods for the ISUAV-D dataset are as follows:

[0018] The ISUAV-D dataset was captured by a long-wave infrared camera, targeting a quadcopter drone against a daytime background of buildings, sky, clouds, pedestrians, and trees, and contains 16 image sequences. The ISUAV-D dataset belongs to the category of datasets with complex background interference. The images are in PNG format with a size of 640×512. The original ISUAV-D dataset contained approximately 20,000 images, including samples where the target disappeared over a long period. These samples were removed in this invention, resulting in a final ISUAV-D dataset containing 13,886 single-target images. The images in the 16 sequences were divided into training, validation, and test sets in an 8:1:1 ratio. A higher-order degradation model was used to simulate degradation of the images in the ISUAV-D dataset in two stages, without using a local degradation process. The degradation factors in the higher-order degradation model were chosen to be relatively small to prevent the target from being completely submerged in the background, resulting in the degraded dataset.

[0019] Furthermore, in S3, the position-aware module performs feature normalization, convolutional mapping, positional attention calculation, positional encoding, scaled dot product calculation, convolution operation, and activation function operation. The input to PAM is the output feature map X of SPPF in YOLOv11. X is first normalized and convolutionally mapped to generate three sets of matrix tensors Q, K, and V for attention calculation. This process can be expressed by the formula:

[0020] Q = W q X,K=W k X+E pos V = W v X

[0021] Among them, E pos The position embedding matrix represents the location encoding, which is embedded in the K vector as external explicit information to compensate for the lack of spatial awareness information in the original Transformer; Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; W q W k W v These represent the query matrix weights, key matrix weights, and value matrix weights, respectively; q, k, and v represent identifiers for the weight matrix functions.

[0022] Attention weights are calculated using a standard scaled dot product, and a region-space mask M is introduced to control the attention focus range. This can be expressed by the following formula:

[0023]

[0024] Among them, F p This represents the output feature map of PAM, Conv(.) represents the convolution operation, Softmax represents the normalized activation function, and ⊙ represents the element-wise multiplication operation of the matrix. Indicates the scaling factor, d is the dimension of the key vector, and the square root controls the size of the dot product; K T This represents the transpose operation of matrix K;

[0025] Finally, the attention output is integrated through convolutional layers to generate enhanced feature maps.

[0026] Furthermore, in S3, the multi-scale saliency fusion module MSS includes a multi-scale layer and a weight generation module; the multi-scale layer consists of convolutional layer one, convolutional layer two, and convolutional layer three, used to generate multi-scale features; the weight generation module consists of a global average pooling layer, a convolutional layer, and a softmax activation function.

[0027] PAM output feature map Simultaneously, through three convolution operations with different kernel sizes, fine-grained F values ​​are extracted respectively. s Medium particle size F m and coarse-grained F l Feature representation, It fully describes a three-dimensional feature map structure. The feature map is defined as having a real number for each pixel value, where H represents the feature map height, W represents the feature map width, and C represents the number of feature map channels. A lightweight, dynamically weighted subnetwork is designed to generate the fusion weights W. s W m and W l The multi-scale saliency fusion feature F is obtained through weighted fusion. d .

[0028] Furthermore, in S3, the position-guided filtering module includes a background estimation network, a target gating layer one, a target gating layer two, and a depth-separable convolutional layer;

[0029] First, the multi-scale saliency fusion feature F d Background response b is generated after background estimation and target gating networks respectively. g and gating factor g t Then obtain the intermediate feature F g This process can be expressed by the formula:

[0030] F g =g t ⊙F p +(1-g t )⊙b g

[0031]

[0032] Where [:] represents a concatenation operation. F represents dPosition encoding, W g Indicates a gating network;

[0033] To capture contrast information in local regions, depthwise separable convolution is used to extract the low-frequency response C. l And construct high-frequency residual C h =F g -C l Finally, the enhanced salient feature F is further modulated by introducing a contrast gating factor. out This process can be expressed by the formula:

[0034] F out =F g +W g (C h )⊙C h .

[0035] Furthermore, in S4, the progressive void rate group strategy establishes spectral-spatial infrared frequency band perception through frequency band adaptive void rate; the void rate group d = [d1, d2, d3] in SSPM is set progressively, with shallow features using a smaller void rate to preserve fine edges and local textures, and deep features using a larger void rate to expand the receptive field and enhance the perception of the overall structure.

[0036] Furthermore, in S4, the channel attention in the local-global dual-path attention module enhances key band features through the SE module with a compression ratio of 0.25, while the spatial attention strengthens the target-background difference through local difference. Finally, entropy weight multi-scale fusion is introduced to achieve adaptive weighting based on the spectral entropy of each branch feature map.

[0037] Furthermore, in step S5, the design method for the sensitivity stabilization loss function DS-IoU is as follows:

[0038] First, define the diagonal length S of the target bounding box. gt As a metric benchmark, S gt Expressed using a formula:

[0039]

[0040] Where w and h represent the width and height of the target bounding box, respectively;

[0041] Then, using the dynamic reference scale D s Construct gradient compensation coefficients γ; D s Expressed using a formula:

[0042] D s =0.1×max(W,H)

[0043] Where W and H represent the width and height of the entire image, respectively;

[0044] The scale penalty weight λ can be expressed by the formula:

[0045]

[0046] The DS-IoU loss function can then be defined as:

[0047]

[0048] Among them, P c This represents the center offset penalty term, which inherits from DIoU; P wh This represents the size penalty; expressed by the formula:

[0049] P c =ρ 2 / c 2

[0050]

[0051] Where ρ represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; c represents the diagonal length of the minimum bounding rectangle of the predicted bounding box and the ground truth bounding box; w and h represent the width and height of the ground truth bounding box, respectively; Δw and Δh represent the absolute differences between the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively.

[0052] The DS-IoU loss function is passed through The penalty scale is dynamically adjusted; when the target size is much smaller than the reference scale, λ decays exponentially, reducing the sensitivity of small target detection loss to positional shifts; aspect ratio differences are also considered. Normalization is applied to eliminate the scale interference of the target's own size on the penalty term; an IoU loss activation factor κ is set to avoid over-optimization under high overlap.

[0053] Furthermore, in S6, the base network is a YOLOv11 network; the reconstruction method refers to replacing the C2PSA and C3k2 modules in the YOLOv11 network with the position-guided enhancement module PSE_Former and the spectral-spatial perception module SSPM to achieve infrared cross-domain feature transfer; the classification loss is the cross-entropy loss.

[0054] (III) Beneficial Effects

[0055] Compared with existing technologies, this invention provides an infrared weak target detection method based on cross-domain feature transfer, which has the following beneficial effects:

[0056] 1. This invention proposes an infrared weak target detection network architecture SA-YOLO, and innovatively designs a position-guided enhancement module and a spectral-spatial sensing module, realizing cross-domain feature transfer from visible light detection network to infrared detection network, improving the spectral-spatial sensing detection performance by about 22%.

[0057] 2. This invention proposes a scale-adaptive loss function DS-IoU, which solves the problems of slow convergence speed and sharp drop in scale sensitivity during the training process of infrared weak target detection network by using dynamic penalty weights and composite normalization.

[0058] 3. Since there are very few publicly available infrared small target detection datasets, this invention constructs the ISUAV-D infrared small target dataset to alleviate the data shortage in this field.

[0059] 4. This invention experimentally verified the superior performance of SA-YOLO in strong light, building occlusion, and small targets, showing competitive results in both detection accuracy and robustness compared to existing methods. Attached Figure Description

[0060] Figure 1 This is a flowchart of an infrared weak target detection method based on cross-domain feature transfer according to the present invention;

[0061] Figure 2 This is an example image of a portion of the samples in the ISUAV-D dataset constructed in this invention;

[0062] Figure 3 A detailed quantitative structure diagram of the ISUAV-D dataset partitioned for this invention;

[0063] Figure 4 The schematic diagram of the position guidance enhancement module PSE_Former designed for this invention is shown below. (a) is a structural diagram of the position awareness module of this invention, and (b) is a structural diagram of the position guidance filtering module of this invention.

[0064] (c) is the multi-scale saliency fusion module of the present invention;

[0065] Figure 5 Schematic diagram of the Spectrum-Spatial Sensing Module (SSPM) designed for this invention;

[0066] Figure 6 This is a conceptual diagram illustrating how the target size analyzed in this invention has a significant impact on IoU sensitivity.

[0067] Figure 7 This is a new SA-YOLO network structure diagram reconstructed according to the present invention;

[0068] Figure 8The figure shows the qualitative comparison results of the present invention with several existing methods on three datasets;

[0069] Figure 9 The figure shows the quantitative comparison results of this invention with several existing methods on three datasets. Detailed Implementation

[0070] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0071] Example:

[0072] like Figure 1 The flowchart shown is a method for detecting weak infrared targets based on cross-domain feature transfer, proposed in one embodiment of the present invention. The method specifically includes the following steps:

[0073] S1. Prepare two public datasets: dataset one is the NUST-SIRST dataset, and dataset two is the SIRST-5k dataset; for details on how to obtain the two datasets and their data structures, please refer to the official documentation.

[0074] The NUST-SIRST dataset contains 10,000 training images, 3,000 validation images, and 1,000 test images.

[0075] The SIRST-5k dataset contains 4862 training images, 300 validation images, and 664 test images.

[0076] S2, Prepare a self-built dataset: including data acquisition, data preprocessing, and data partitioning;

[0077] The self-built dataset of this invention is named ISUAV-D, as follows: Figure 2 The image shown is a sample of some ISUAV-D images. ISUAV-D was captured by a long-wave infrared camera, targeting a drone, and consists of 16 image sequences with a daytime background including buildings, sky, clouds, pedestrians, and trees. ISUAV-D belongs to the complex background interference class dataset. The image format is PNG, and the image size is 640×512.

[0078] The original ISUAV-D contained approximately 20,000 images, including samples where the target disappeared over a long period. These samples were removed in this invention, resulting in a final ISUAV-D set of 13,886 single-target images. This invention divides the images from 16 sequences into training, validation, and test sets in an 8:1:1 ratio. Figure 3 The image shows detailed information about the ISUAV-D partitioning. This invention uses a higher-order degradation model to perform a two-stage simulated degradation of ISUAV-D (without using a local degradation process). The degradation factors of the higher-order degradation model should be selected to be relatively small to prevent the target from being completely submerged in the background, and finally obtain the degraded dataset.

[0079] S3, Design the Position Guidance Enhancement Module PSE_Former: PSE_Former includes the Position Awareness Module (PAM), the Multi-Scale Saliency Fusion Module (MSS), and the Position Guidance Filtering Module (PGG); PAM, MSS, and PGG are cascaded within PSE_Former; the overall structure of PSE_Former is as follows: Figure 4 As shown;

[0080] Design of Position Aware Module (PAM): such as Figure 4 As shown in (a), the Position Awareness Module (PAM) performs feature normalization, convolutional mapping, positional attention calculation, positional encoding, scaled dot product calculation, convolution operation, and activation function operation. The input to PAM is the output feature map X of SPPF in YOLOv11. X is first normalized and convolutionally mapped to generate three sets of matrix tensors Q, K, and V for attention calculation. This process can be expressed by the following formula:

[0081] Q = W q X,K=W k X+E pos V = W v X

[0082] Among them, E pos The position embedding matrix represents the location encoding, which is embedded in the K vector as external explicit information to compensate for the lack of spatial awareness information in the original Transformer; Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; W q W k W v These represent the query matrix weights, key matrix weights, and value matrix weights, respectively; q, k, and v represent identifiers for the weight matrix functions.

[0083] Attention weights are calculated using the standard scaled dot product. This paper introduces a region-space mask M to control the attention focus range, which can be expressed by the formula:

[0084]

[0085] Among them, F p This represents the output feature map of PAM, Conv(.) represents the convolution operation, Softmax represents the normalized activation function, and ⊙ represents the element-wise multiplication operation of the matrix. Indicates the scaling factor, d is the dimension of the key vector, and the square root controls the size of the dot product; K T This represents the transpose operation of matrix K;

[0086] Finally, the attention output is integrated through convolutional layers to generate enhanced feature maps;

[0087] Design of the Multi-Scale Saliency Fusion Module (MSS): Since infrared weak targets often exhibit structural changes at different scales, this invention utilizes MSS to extract target features under different receptive fields and employs pyramid fusion with dynamic weight allocation to achieve adaptive fusion; for example... Figure 4 As shown in (c), the Multi-Scale Saliency Fusion Module (MSS) includes a multi-scale layer and a weight generation module. The multi-scale layer consists of convolutional layer 1, convolutional layer 2, and convolutional layer 3, which are used to generate multi-scale features. The weight generation module consists of a global average pooling layer, a convolutional layer, and a Softmax activation function.

[0088] PAM output feature map Simultaneously, through three convolution operations with different kernel sizes, fine-grained F values ​​are extracted respectively. s Medium particle size F m and coarse-grained F l Feature representation, It fully describes a three-dimensional feature map structure. The feature map is defined as having a real number for each pixel value, where H represents the feature map height, W represents the feature map width, and C represents the number of feature map channels. A lightweight, dynamically weighted subnetwork is designed to generate the fusion weights W. s W m and W l The multi-scale saliency fusion feature F is obtained through weighted fusion. d .

[0089] Design of the Position-Guided Filtering Module (PGG): Based on the prior knowledge that weak infrared targets are mostly located in specific frequency bands, this invention utilizes a gating mechanism to achieve adaptive background elimination; such as... Figure 4 As shown in (b), the Position Guided Filtering (PGG) module includes a background estimation network, a target gating layer 1, a target gating layer 2, and a depthwise separable convolutional layer.

[0090] First, the multi-scale saliency fusion feature F d Background response b is generated after background estimation and target gating networks respectively. g and gating factor gt Then obtain the intermediate feature F g This process can be expressed by the formula:

[0091] F g =g t ⊙F p +(1-g t )⊙b g

[0092]

[0093] Where [:] represents a concatenation operation. F represents d Position encoding, W g Indicates a gating network;

[0094] To capture contrast information in local regions, this invention utilizes depthwise separable convolution to extract the low-frequency response C. l And construct high-frequency residual C h =F g -C l Finally, the enhanced salient feature F is further modulated by introducing a contrast gating factor. out This process can be expressed by the formula:

[0095] F out =F g +W g (C h )⊙C h

[0096] S4, Design Spectrum-Spatial Awareness Module SSPM: SSPM includes a progressive void rate grouping strategy and a local-global dual-path attention module LGAM;

[0097] Progressive Hole Rate Grouping Strategy: The original YOLOv11's C3k2 module, by introducing multi-scale convolutional kernels based on the C3 module, is more suitable for detecting large objects in visible light images but not for detecting small infrared targets. Therefore, this invention designs SSPM, which establishes spectral-spatial infrared band sensing through frequency band adaptive hole rate. The structure of SSPM is as follows: Figure 5 As shown; the progressive void ratio group d = [d1, d2, d3] in SSPM is set as follows: Figure 7 As shown, shallow features use a smaller void ratio to preserve fine edges and local textures, while deep features use a larger void ratio to expand the receptive field and enhance the perception of the overall structure.

[0098] Local-Global Dual Attention Module LGAM: In SSPM, this invention designs a local-global dual attention module, in which channel attention enhances key band features through an SE module with a compression ratio of 0.25, while spatial attention strengthens the target-background difference through local difference. Finally, entropy weight multi-scale fusion is introduced, and adaptive weighting is achieved based on the spectral entropy of each branch feature map.

[0099] S5, Design the sensitivity-stabilized loss function DS-IoU: DS-IoU will be trained in conjunction with ordinary IoU to train the network model. The weights of DS-IoU and ordinary IoU are determined by the loss activation factor κ, which is a hyperparameter.

[0100] Problem analysis: such as Figure 6 As shown, the target size has a significant impact on IoU sensitivity; for a small target of 6×6 pixels, the IoU value drops sharply from 0.53 to 0.06 (a decrease of 88.7%) when the diagonal of the prediction box is shifted by 1 pixel, while for a regular target of 24×24 pixels, the IoU drops from 0.75 to 0.29 (a decrease of 37.6%) under the same shift; GIOU

[39] and CIOU

[40] have the same phenomenon; in order to suppress this effect, this invention proposes a sensitivity-stable loss function DS-IoU, which achieves robust optimization of target scale differences through dynamic penalty weights and composite normalization mechanism;

[0101] Design method: First, define the diagonal length S of the target bounding box. gt As a metric benchmark, S gt Expressed using a formula:

[0102]

[0103] Where w and h represent the width and height of the target bounding box, respectively;

[0104] Then, using the dynamic reference scale D s Construct the gradient compensation coefficient γ. s Expressed using a formula:

[0105] D s =0.1×max(W,H)

[0106] Where W and H represent the width and height of the entire image, respectively;

[0107] The scale penalty weight λ can be expressed by the formula:

[0108]

[0109] The DS-IoU loss function can then be defined as:

[0110]

[0111] Among them, P c This represents the center offset penalty term, which inherits from DIoU; P wh This represents the size penalty; expressed by the formula:

[0112] P c =ρ 2 / c 2

[0113]

[0114] Where ρ represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; c represents the diagonal length of the minimum bounding rectangle of the predicted bounding box and the ground truth bounding box; w and h represent the width and height of the ground truth bounding box, respectively; Δw and Δh represent the absolute differences between the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively.

[0115] This invention is achieved through The penalty scale is dynamically adjusted; when the target size is much smaller than the reference scale, λ decays exponentially, reducing the sensitivity of small target detection loss to positional shifts; aspect ratio differences are also considered. Normalization is applied to eliminate the scale interference of the target's own size on the penalty term; an IoU loss activation factor κ is set to avoid over-optimization under high overlap.

[0116] S6, Network Reconstruction and Model Training: Using PSE_Former from step S3 and SSPM from step S4, reconstruct the new network SA-YOLO on the base YOLO network, as follows: Figure 7 As shown; the model is trained using the training sets of the three datasets in steps S1 and S2, and the DS-IoU loss and classification loss in step S5. At the same time, the model is fine-tuned using the validation sets of the three datasets to further improve the cross-domain feature transfer detection performance from visible light to infrared.

[0117] The base network is the YOLOv11 network; the reconstruction method refers to replacing the C2PSA and C3k2 modules in the YOLOv11 network with the position-guided enhancement module PSE_Former and the spectral-spatial perception module SSPM to achieve infrared cross-domain feature transfer; the classification loss is the cross-entropy loss;

[0118] S7, Model Testing and Evaluation: After the fine-tuning in step S6 is completed, solidify the fine-tuned network parameters and save the final weak target detection model; use the test set in steps S1 and S2 to test the model effect from a qualitative and quantitative perspective. Real infrared images can be directly input into the trained end-to-end model to obtain result images labeled with detection boxes and location information.

[0119] Qualitative evaluation refers to visualizing the detection results of images; quantitative evaluation refers to evaluating the performance of the model through detection evaluation metrics. Quantitative evaluation metrics include prediction accuracy AP3p25, detection accuracy Acc at recall = 0.9, average accuracy mAP50 at intersection-to-union ratio greater than 50, and average accuracy mAP50-95 at intersection-to-union ratio greater than 50 and less than 95.

[0120] This invention proposes a novel spectral sensing network, SA-YOLO, to address the key challenge that YOLO models trained in visible light scenarios cannot be directly applied to infrared small target detection tasks. First, the invention designs the PSE_Former module, which significantly enhances the spectral feature representation capability of infrared small targets through a triple structure of PAM, MSS, and PGG, solving the problem of feature overload in traditional methods under low signal-to-noise ratio conditions. Second, the proposed SSPM employs a progressive hole rate and a dual-branch attention mechanism to achieve robust extraction of multi-scale features. Then, addressing the scale sensitivity of infrared targets, an innovative scale-adaptive DS-IoU loss is proposed, which reduces the localization error of targets of different sizes through dynamic penalty weights and a composite normalization mechanism. Finally, this invention constructs the ISUAV-D infrared UAV small target dataset. Extensive experimental verification shows that SA-YOLO exhibits significant advantages in environments with strong light, clouds, and occlusion interference, such as… Figure 8 and Figure 9 As shown. In the future, this invention plans to extend SA-YOLO to a wider range of infrared sensing tasks, such as multispectral collaborative detection and dynamic target tracking.

[0121] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting weak infrared targets based on cross-domain feature transfer, characterized in that, Includes the following steps: S1. Prepare two public datasets: dataset one is the NUST-SIRST dataset, and dataset two is the SIRST-5k dataset. S2. Prepare a self-built dataset, including data acquisition, data preprocessing, and data partitioning; name the self-built dataset ISUAV-D. S3, Design the Position Guidance Enhancement Module PSE_Former: PSE_Former includes the Position Awareness Module PAM, the Multi-Scale Saliency Fusion Module MSS, and the Position Guidance Filtering Module PGG; PAM, MSS, and PGG are cascaded in PSE_Former; S4, Design Spectrum-Spatial Awareness Module SSPM: SSPM includes a progressive void rate grouping strategy and a local-global dual-path attention module LGAM; S5, Design the sensitivity-stabilized loss function DS-IoU: DS-IoU will be trained in conjunction with ordinary IoU to train the network model. The weights of DS-IoU and ordinary IoU are determined by the loss activation factor κ, which is a hyperparameter. S6, Network Reconstruction and Model Training: Reconstruct a new network SA-YOLO on the base YOLO network using PSE_Former from step S3 and SSPM from step S4; Train the model using the training sets of the three datasets from steps S1 and S2, and the DS-IoU loss and classification loss from step S5, while fine-tuning the model using the validation sets of the three datasets. In step S6, the base network is a YOLOv11 network; the reconstruction method refers to replacing the C2PSA and C3k2 modules in the YOLOv11 network with the position-guided enhancement module PSE_Former and the spectral-spatial perception module SSPM to achieve infrared cross-domain feature transfer; the classification loss is cross-entropy loss. S7, Model Testing and Evaluation: After the fine-tuning in step S6 is completed, solidify the fine-tuned network parameters and save the final weak target detection model; use the test set in steps S1 and S2 to test the model effect from a qualitative and quantitative perspective. Real infrared images can be directly input into the trained end-to-end model to obtain result images labeled with detection boxes and location information.

2. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: For details regarding the acquisition method and data structure of the NUST-SIRST and SIRST-5k datasets in S1, please refer to the existing documentation.

3. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In step S2, the data acquisition, data preprocessing, and data partitioning methods for the ISUAV-D dataset are as follows: The ISUAV-D dataset was captured by a long-wave infrared camera, targeting a quadcopter drone against a daytime background of buildings, sky, clouds, pedestrians, and trees, and contains 16 image sequences. The ISUAV-D dataset belongs to the category of datasets with complex background interference. The images are in PNG format with a size of 640×512. The original ISUAV-D dataset contained approximately 20,000 images, including samples where the target disappeared over a long period. These samples were removed in this invention, resulting in a final ISUAV-D dataset containing 13,886 single-target images. The images in the 16 sequences were divided into training, validation, and test sets in an 8:1:1 ratio. A higher-order degradation model was used to simulate degradation of the images in the ISUAV-D dataset in two stages, without using a local degradation process. The degradation factors in the higher-order degradation model were chosen to be relatively small to prevent the target from being completely submerged in the background, resulting in the degraded dataset.

4. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In step S3, the position-aware module performs feature normalization, convolutional mapping, positional attention calculation, positional encoding, scaled dot product calculation, convolution operation, and activation function operation. The input to PAM is the output feature map X of SPPF in YOLOv11. X is first normalized and convolutionally mapped to generate three sets of matrix tensors Q, K, and V for attention calculation. This process can be expressed by the following formula: Q=W q X,K=W k X+E pos ,V=W v X Among them, E pos The position embedding matrix represents the location encoding, which is embedded in the K vector as external explicit information to compensate for the lack of spatial awareness information in the original Transformer; Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; W q W k W v These represent the query matrix weights, key matrix weights, and value matrix weights, respectively; q, k, and v represent identifiers for the weight matrix functions. Attention weights are calculated using a standard scaled dot product, and a region-space mask M is introduced to control the attention focus range. This can be expressed by the following formula: Among them, F p This represents the output feature map of PAM, Conv(.) represents the convolution operation, Softmax represents the normalized activation function, and ⊙ represents the element-wise multiplication operation of the matrix. The scaling factor is represented by d, which represents the dimension of the key vector. The square root controls the size of the dot product. T This represents the transpose operation of matrix K; Finally, the attention output is integrated through convolutional layers to generate enhanced feature maps.

5. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In S3, the multi-scale saliency fusion module MSS includes a multi-scale layer and a weight generation module; the multi-scale layer consists of convolutional layer one, convolutional layer two and convolutional layer three, and is used to generate multi-scale features; The weight generation module consists of a global average pooling layer, a convolutional layer, and a softmax activation function; PAM output feature map Simultaneously, through three convolution operations with different kernel sizes, fine-grained F values ​​are extracted respectively. s Medium particle size F m and coarse-grained F l Feature representation, It fully describes a three-dimensional feature map structure. The feature map is defined as follows: each pixel value is a real number; H represents the feature map height; W represents the feature map width; and C represents the number of feature map channels. A lightweight, dynamically weighted sub-network is designed to generate the fusion weights W. s W m and W l The multi-scale saliency fusion feature F is obtained through weighted fusion. d .

6. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In S3, the position-guided filtering module includes a background estimation network, a target gating layer one, a target gating layer two, and a depth-separable convolutional layer. First, the multi-scale saliency fusion feature F d Background response b is generated after background estimation and target gating networks respectively. g and gating factor g t Then obtain the intermediate feature F g This process can be expressed by the formula: F g =g t ⊙F p +(1-g t )⊙b g Where [:] represents a concatenation operation. F represents d Position encoding, W g Indicates a gating network; To capture contrast information in local regions, depthwise separable convolution is used to extract the low-frequency response C. l And construct high-frequency residual C h =F g -C l Finally, the enhanced salient feature F is further modulated by introducing a contrast gating factor. out This process can be expressed by the formula: F out =F g +W g (C h )⊙C h 。 7. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In S4, the progressive void rate group strategy establishes spectral-spatial infrared frequency band perception through frequency band adaptive void rate; the void rate group d = [d1, d2, d3] in SSPM is set progressively, shallow features use a smaller void rate to preserve fine edges and local textures, and deep features use a larger void rate to expand the receptive field and enhance the perception of the overall structure.

8. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In S4, the channel attention in the local-global dual-path attention module enhances key band features through the SE module with a compression ratio of 0.25, while the spatial attention strengthens the target-background difference through local difference. Finally, entropy weight multi-scale fusion is introduced, and adaptive weighting is achieved based on the spectral entropy of each branch feature map.

9. The infrared weak target detection method based on cross-domain feature transfer according to claim 1, characterized in that: In step S5, the design method for the sensitivity-stabilized loss function DS-IoU is as follows: First, define the diagonal length S of the target bounding box. gt As a metric benchmark, S gt Expressed using a formula: Where w and h represent the width and height of the target bounding box, respectively; Then, using the dynamic reference scale D s Construct gradient compensation coefficients γ; D s Expressed using a formula: D s =0.1×max(W,H) Where W and H represent the width and height of the entire image, respectively; The scale penalty weight λ can be expressed by the formula: The DS-IoU loss function can then be defined as: Among them, P c This represents the center offset penalty term, which inherits from DIoU; P wh This represents the size penalty; expressed by the formula: P.S c Zρ 2 / c 2 Where ρ represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; c represents the diagonal length of the minimum bounding rectangle of the predicted bounding box and the ground truth bounding box; w and h represent the width and height of the ground truth bounding box, respectively; Δw and Δh represent the absolute differences between the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively. The DS-IoU loss function is passed through The penalty scale is dynamically adjusted; when the target size is much smaller than the reference scale, λ decays exponentially, reducing the sensitivity of small target detection loss to positional shifts; aspect ratio differences are also considered. Normalization is applied to eliminate the scale interference of the target's own size on the penalty term; an IoU loss activation factor κ is set to avoid over-optimization under high overlap.

Images