Infrared weak and small target detection method based on cross-domain feature migration

Abstract

The invention belongs to the technical field of target detection, particularly relates to an infrared weak and small target detection method based on cross-domain feature migration, and provides a novel detection network which is SA-YOLO for short. Firstly, a position guidance enhancement module is designed under the inspiration of a multispectral feature migration idea, and infrared spectral feature response of a weak and small target is effectively enhanced through a triple structure of position perception attention, multi-scale significance fusion and position guidance filtering. Secondly, a spectrum-space sensing module is provided, and a progressive voidage and a double-branch attention mechanism are adopted, so that robustness modeling of multi-scale spatial features is realized. Then, a sensitivity stability loss function is innovatively proposed, and the problem of sudden drop of scale sensitivity of a weak and small target is relieved through dynamic penalty weight and composite normalization. And finally, an infrared small target data set named as ISUAV-D is also collected. Experimental results show that SA-YOLO shows significant advantages in a strong light interference environment and a tiny target detection scene.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to an infrared dim small target detection method based on cross-domain feature migration. Background Art

[0002] In complex environments such as military reconnaissance and security monitoring, infrared imaging systems can effectively overcome the limitations of visible light sensors and achieve stable detection of long-range, low signal-to-noise ratio targets. However, small infrared targets are usually composed of only a small number of pixels, lack obvious shape and texture features, and are easily interfered by complex backgrounds such as clouds and strong light. Existing methods are difficult to achieve high-precision detection that exceeds that of large visible light targets. Therefore, finding a method that can effectively transfer visible light training pre-networks to infrared detection tasks is an urgent problem to be solved. This paper mainly studies how to overcome the domain offset problem between visible light and infrared bands while retaining the powerful feature extraction ability of pre-trained networks through modular design of infrared spectral characteristics.

[0003] To solve the above problems, traditional infrared dim target detection technology is mainly based on manually designed features and prior knowledge. This type of technology performs well under uniform backgrounds, but it relies on complex matrix operations and has huge computational overhead when processing high-resolution images. Thanks to deep learning and large amounts of data, detection based on convolutional neural networks has made many breakthroughs. Mainstream technologies include single-stage detectors (YOLO series) and two-stage detectors. Although the existing single-stage detectors and two-stage detectors have improved the network's feature extraction capabilities by introducing attention mechanisms and other means, they have not studied the problem of infrared cross-domain feature migration. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides an infrared dim target detection method based on cross-domain feature migration, which solves the problem of low detection accuracy caused by the existing direct use of visible light detection models due to challenges such as low signal-to-noise ratio and strong background interference of infrared dim targets; and solves the problem of sudden drop in scale sensitivity of infrared dim targets in the traditional IoU loss function when training regression algorithms.

[0006] (2) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0008] A method for detecting infrared dim small targets based on cross-domain feature migration specifically includes the following steps:

[0009] S1, prepare two public datasets: Dataset 1 and Dataset 2 both contain training sets, validation sets, and test sets;

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

[0011] S3, design the position-guided enhancement module PSE_Former: PSE_Former includes the position-aware module PAM, the multi-scale saliency fusion module MSS, and the position-guided filtering module PGG; PAM, MSS, and PGG are in a cascade relationship in PSE_Former;

[0012] S4, design of spectral-spatial perception module SSPM: SSPM includes progressive dilation rate group strategy and local-global dual-path attention module LGAM;

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

[0014] S6, network reconstruction and model training: Use the PSE_Former in step S3 and the SSPM in step S4 to reconstruct the new network SA-YOLO on the basic YOLO network; use the training sets of the three datasets in steps S1 and S2, and the DS-IoU loss and classification loss in step S5 to train the model, and use the validation sets of the three datasets to fine-tune the model to further improve the cross-domain feature migration 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 small target detection model; use the test sets in steps S1 and S2 to test the model effect from qualitative and quantitative perspectives. The real infrared image can be directly input into the trained end-to-end model to obtain the result image marked with detection box and position information.

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

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

[0018] The ISUAV-D dataset was taken by a long-wave infrared camera. The target is a quadrotor drone, and the background is buildings, sky, white clouds, pedestrians, and trees in a daytime scene. It contains 16 image sequences. The ISUAV-D dataset belongs to the complex background interference category. The image format is PNG format. The image size is 640×512. The total number of images in the original ISUAV-D dataset is 20,000, which includes samples in which the target disappears for a long time. These samples were removed, and the final ISUAV-D dataset contains 13,886 single-target images. The images in the 16 sequences are divided into training set, validation set, and test set in an 8:1:1 ratio. A high-order degradation model is used to perform two-stage simulated degradation on the images in the ISUAV-D dataset. No local degradation process is used. The degradation factors of the high-order degradation model should be selected to be smaller to prevent the target from being completely submerged in the background, and finally a degraded dataset is obtained.

[0019] Furthermore, in S3, the position perception module performs feature normalization, convolution mapping, position attention calculation, position encoding, scaled dot product calculation, convolution operation and activation function operation; the input of PAM is the output feature map of SPPF in YOLOv11 , First, after normalization and convolution mapping, three sets of matrix tensors Q, K, and V are generated for attention calculation. This process can be expressed as follows:

[0020]

[0021] in, Represents the position embedding matrix. The position encoding is embedded in the K vector as external explicit information to compensate for the lack of spatial perception information of the original Transformer. Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. 、 、 Respectively represent the query matrix weight, key matrix weight and value matrix weight; q, k, v represent the identification of the weight matrix function;

[0022] The attention weight is calculated using the standard scaled dot product, introducing the regional spatial mask Controlling the focus range can be expressed using the formula:

[0023]

[0024] in, represents the output feature map of PAM, Represents the convolution operation, Softmax represents the normalized activation function, Represents the matrix corresponding element multiplication operation, represents the scaling factor, d is the dimension of the key vector, and the square root is used to control the size of the dot product; Representation matrix The transpose operation;

[0025] Finally, the attention outputs are integrated through the convolutional layer 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 is composed of convolutional layer 1, convolutional layer 2 and convolutional layer 3, and is used to generate multi-scale features; the weight generation module is composed of a global average pooling layer, a convolutional layer and a softmax activation function;

[0027] Output feature map of PAM At the same time, three convolution operations with different convolution kernel sizes are performed to extract fine-grained , medium particle size and coarse-grained Feature representation, Completely describes a three-dimensional feature graph structure, Indicates that each pixel value in the feature map 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 dynamic weighted subnetwork is designed to generate fusion weights , and , multi-scale saliency fusion features are obtained through weighted fusion .

[0028] Furthermore, in S3, the position-guided filtering module includes a background estimation network, a target gating layer 1, a target gating layer 2, and a depth-separable convolutional layer;

[0029] First, multi-scale saliency fusion features Generate background response through background estimation and target gating network respectively and gating factors , and then get the intermediate features , this process can be expressed as:

[0030]

[0031]

[0032] in, Represents a splicing operation, express Positional encoding, represents the gating network;

[0033] In order to capture the contrast information of local areas, depthwise separable convolution is used to extract low-frequency responses. , and construct high-frequency residuals Finally, the enhanced salient features are further adjusted by introducing the contrast gating factor , this process can be expressed as:

[0034] .

[0035] Furthermore, in the S4, the progressive hole rate group strategy establishes spectrum-spatial infrared band perception through frequency band adaptive hole rate; the hole rate group in SSPM The setting is progressive. The shallow features use a small void ratio to retain fine edges and local textures, and the deep features use a large void ratio 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 the key band features through the SE module with a compression ratio of 0.25, and the spatial attention strengthens the target-background difference through local difference. Finally, entropy weighted multi-scale fusion is introduced to achieve adaptive weighting based on the spectral entropy of each branch feature map.

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

[0038] First, define the diagonal length of the target box As a benchmark, The formula is:

[0039]

[0040] in, and Represent the width and height of the target box respectively;

[0041] Then, using the dynamic benchmark scale Constructing gradient compensation coefficients ; The formula is:

[0042]

[0043] in, and Respectively represent the width and height of the entire image;

[0044] The scale penalty weight It can be expressed using the formula:

[0045]

[0046] Then the DS-IoU loss function can be defined as:

[0047]

[0048] in, represents the center offset penalty term, which inherits DIoU; Represents the size penalty term; it is expressed as follows:

[0049]

[0050]

[0051] in, Represents the Euclidean distance between the center point of the predicted box and the center point of the real box; Represents the diagonal length of the minimum enclosing rectangle of the predicted box and the true box; and Represent the width and height of the real box respectively; and Represents the absolute difference between the width and height of the predicted box and the width and height of the real box respectively;

[0052] The DS-IoU loss function is obtained by Dynamically adjust the penalty scale. When the target size is smaller than the benchmark scale, Exponential decay makes the small target detection loss less sensitive to position offset; Normalization processing to eliminate the scale interference of the target size on the penalty term; set the IoU loss activation factor , to avoid over-optimization under high overlap.

[0053] Furthermore, in S6, the basic YOLO network is a YOLOv11 network; the reconstruction method refers to: using a position-guided enhancement module PSE_Former and a spectral-spatial perception module SSPM to replace the C2PSA module and the C3k2 module in the YOLOv11 network to achieve infrared cross-domain feature migration; the classification loss is a cross-entropy loss.

[0054] (3) Beneficial effects

[0055] Compared with the existing technology, the present invention provides an infrared dim small target detection method based on cross-domain feature migration, which has the following beneficial effects:

[0056] 1. This paper proposes an infrared small target detection network architecture SA-YOLO, innovatively designs a position guidance enhancement module and a spectral-spatial perception module, realizes cross-domain feature migration from the visible light detection network to the infrared detection network, and improves the spectral-spatial perception detection performance by about 22%.

[0057] 2. This paper proposes a scale-adaptive loss function DS-IoU, which solves the problems of slow convergence and sudden drop in scale sensitivity during the training process of infrared dim target detection network through dynamic penalty weights and composite normalization.

[0058] 3. Since there are currently few public infrared small target detection datasets, this paper constructs the ISUAV-D infrared small target dataset to alleviate the data dilemma in this field.

[0059] 4. This paper experimentally verifies the superior performance of SA-YOLO in strong light, building occlusion, and small targets, and shows competitive results in detection accuracy and robustness compared to existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of an infrared dim small target detection method based on cross-domain feature migration of the present invention;

[0061] Figure 2 This is a sample diagram of a portion of the ISUAV-D dataset constructed by the present invention;

[0062] Figure 3 Detailed quantitative structure diagram of the ISUAV-D dataset constructed for the present invention;

[0063] Figure 4 Schematic diagram of the position-guided enhancement module PSE_Former designed for the present invention, where (a) is the structure diagram of the position perception module of the present invention, (b) is the structure diagram of the position-guided filtering module of the present invention, and (c) is the multi-scale saliency fusion module of the present invention;

[0064] Figure 5 Schematic diagram of the spectral-spatial perception module (SSPM) designed for the present invention;

[0065] Figure 6 Concept diagram showing that the object size has a significant impact on IoU sensitivity as analyzed by the present invention;

[0066] Figure 7 A new SA-YOLO network structure diagram reconstructed by the present invention;

[0067] Figure 8 The figure shows the qualitative comparison results of the present invention with several existing methods on three data sets;

[0068] Figure 9 The figure shows the quantitative comparison results of the present invention with several existing methods on three data sets. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0070] Example:

[0071] like Figure 1 As shown in FIG, a flowchart of a method for detecting infrared dim small targets based on cross-domain feature migration proposed in one embodiment of the present invention, the method specifically includes the following steps:

[0072] S1. Prepare two public datasets: the first dataset is the NUST-SIRST dataset, and the second dataset is the SIRST-5k dataset. The specific acquisition methods and data structures of the two datasets can be found in the official instructions.

[0073] The NUST-SIRST dataset contains 10,000 training images, 3,000 validation images, and 1,000 test images;

[0074] The SIRST-5k dataset contains 4862 training images, 300 validation images, and 664 test images;

[0075] S2, prepare a self-built dataset: including data acquisition, data preprocessing and data partitioning;

[0076] The self-built dataset of the present invention is named ISUAV-D. Figure 2 The following are some sample examples from ISUAV-D. ISUAV-D is captured by a long-wave infrared camera. The target is a drone, and the background is a daytime scene with 16 image sequences of buildings, sky, white clouds, pedestrians, trees, etc. ISUAV-D belongs to the complex background interference dataset. The image format is PNG and the image size is 640×512.

[0077] The total number of images in the original ISUAV-D is 20,000, including samples where the target disappears for a long time. The present invention removes these samples. The final ISUAV-D contains 13,886 single-target images. The images in the 16 sequences are divided into training set, validation set and test set according to the ratio of 8:1:1. Figure 3The figure shows the detailed information of the ISUAV-D partitioning. The ISUAV-D is degraded in two stages using a high-order degradation model (without using a local degradation process). The degradation factors of the high-order degradation model should be small to prevent the target from being completely submerged in the background. The degraded dataset is finally obtained.

[0078] S3, design the position-guided enhancement module PSE_Former: PSE_Former includes the position-aware module PAM, the multi-scale saliency fusion module MSS and the position-guided filtering module PGG; PAM, MSS and PGG are in a cascade relationship in PSE_Former; the overall structure of PSE_Former is as follows Figure 4 As shown;

[0079] Design of Position Awareness Module PAM: Figure 4 As shown in (a), the position perception module PAM performs feature normalization, convolution mapping, position attention calculation, position encoding, scaled dot product calculation, convolution operation and activation function operation; the input of PAM is the output feature map of SPPF in YOLOv11 , First, after normalization and convolution mapping, three sets of matrix tensors Q, K, and V are generated for attention calculation. This process can be expressed as follows:

[0080]

[0081] in, Represents the position embedding matrix. The position encoding is embedded in the K vector as external explicit information to compensate for the lack of spatial perception information of the original Transformer. Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. 、 、 Respectively represent the query matrix weight, key matrix weight and value matrix weight; q, k, v represent the identification of the weight matrix function;

[0082] The attention weight is calculated using the standard scaled dot product. This paper introduces the regional spatial mask Controlling the focus range can be expressed using the formula:

[0083]

[0084] in, represents the output feature map of PAM, Represents the convolution operation, Softmax represents the normalized activation function, Represents the matrix corresponding element multiplication operation, represents the scaling factor, d is the dimension of the key vector, and the square root is used to control the size of the dot product; Representation matrix The transpose operation;

[0085] Finally, the attention output is integrated through the convolution layer to generate an enhanced feature map;

[0086] Design of Multi-Scale Saliency Fusion Module MSS: Since infrared weak targets often have structural changes at different scales, this paper uses MSS to extract target features under different receptive fields and uses pyramid fusion with dynamic weight allocation to achieve adaptive fusion; Figure 4 As shown in (c), the multi-scale saliency fusion module MSS contains 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 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;

[0087] Output feature map of PAM At the same time, three convolution operations with different convolution kernel sizes are performed to extract fine-grained , medium particle size and coarse-grained Feature representation, Completely describes a three-dimensional feature graph structure, Indicates that each pixel value in the feature map 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 dynamic weighted subnetwork is designed to generate fusion weights , and , multi-scale saliency fusion features are obtained through weighted fusion .

[0088] Design of the position-guided filtering module PGG: Due to the prior knowledge that most weak infrared targets are located in a specific frequency band, the present invention uses a gating mechanism to achieve adaptive background elimination; Figure 4 As shown in (b), the position-guided filtering module PGG includes a background estimation network, a target gating layer 1, a target gating layer 2, and a depth-separable convolutional layer;

[0089] First, multi-scale saliency fusion features Generate background response through background estimation and target gating network respectively and gating factors , and then get the intermediate features , this process can be expressed as:

[0090]

[0091]

[0092] in, Represents a splicing operation, express Positional encoding, represents the gating network;

[0093] In order to capture the contrast information of local areas, the present invention uses depthwise separable convolution to extract low-frequency response. , and construct high-frequency residuals Finally, the enhanced salient features are further adjusted by introducing the contrast gating factor , this process can be expressed as:

[0094]

[0095] S4, design of spectral-spatial perception module SSPM: SSPM includes progressive dilation rate group strategy and local-global dual-path attention module LGAM;

[0096] Progressive Hollowness Group Strategy: The C3k2 module in the original YOLOv11 introduces a multi-scale convolution kernel based on the C3 module, which is more suitable for the detection of large objects in visible light images, but not for the detection of weak infrared targets. Therefore, the present invention designs SSPM, which establishes spectral-spatial infrared band perception through frequency band adaptive holeness rate. The structure of SSPM is as follows: Figure 5 As shown; Progressive void rate group in SSPM The settings are as follows Figure 7 As shown in the figure, shallow features use a small dilation rate to retain fine edges and local textures, while deep features use a large dilation rate to expand the receptive field and enhance the perception of the overall structure.

[0097] Local-global dual-path attention module (LGAM): In SSPM, the present invention designs a local-global dual-path attention, in which the channel attention enhances the key band features through the SE module with a compression ratio of 0.25, and the spatial attention strengthens the target-background difference through local difference. Finally, entropy weighted multi-scale fusion is introduced to achieve adaptive weighting based on the spectral entropy of each branch feature map.

[0098] S5, design sensitivity stability loss function DS-IoU: The sensitivity stability loss function DS-IoU includes ordinary IoU and DS-IoU, and the two parts are composed of loss activation factors Decide, is a hyperparameter;

[0099] Problem Analysis: If Figure 6As shown in the figure, the target size has a significant impact on the sensitivity of IoU. For a small target of 6×6 pixels, a 1-pixel diagonal offset of the prediction box causes the IoU value to drop sharply from 0.53 to 0.06 (a decrease of 88.7%), 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 offset. GIOU

[39] and CIOU

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

[0100] Design method: First, define the diagonal length of the target box As a benchmark, The formula is:

[0101]

[0102] in, and Represent the width and height of the target box respectively;

[0103] Then, using the dynamic benchmark scale Constructing gradient compensation coefficients . The formula is:

[0104]

[0105] in, and Respectively represent the width and height of the entire image;

[0106] The scale penalty weight It can be expressed using the formula:

[0107]

[0108] Then the DS-IoU loss function can be defined as:

[0109]

[0110] in, represents the center offset penalty term, which inherits DIoU; Represents the size penalty term; it is expressed as follows:

[0111]

[0112]

[0113] in, Represents the Euclidean distance between the center point of the predicted box and the center point of the real box; Represents the diagonal length of the minimum enclosing rectangle of the predicted box and the true box; and Represent the width and height of the real box respectively; and Represents the absolute difference between the width and height of the predicted box and the width and height of the real box respectively;

[0114] The present invention Dynamically adjust the penalty scale. When the target size is smaller than the benchmark scale, Exponential decay makes the small target detection loss less sensitive to position offset; Normalization processing to eliminate the scale interference of the target size on the penalty term; set the IoU loss activation factor , avoid over-optimization under high overlap;

[0115] S6, network reconstruction and model training: Use the PSE_Former of step S3 and the SSPM of step S4 to reconstruct the new network SA-YOLO on the basic YOLO network, such as Figure 7 As shown in the figure, 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 migration detection performance from visible light to infrared.

[0116] The basic YOLO network is the YOLOv11 network. The reconstruction method is to use the position-guided enhancement module PSE_Former and the spectral-spatial perception module SSPM to replace the C2PSA module and C3k2 module in the YOLOv11 network to achieve infrared cross-domain feature migration. The classification loss is the cross-entropy loss.

[0117] S7, model testing and evaluation: After the fine-tuning in step S6 is completed, solidify the fine-tuned network parameters and save the final small target detection model; use the test sets in steps S1 and S2 to test the model effect from qualitative and quantitative perspectives. The real infrared image can be directly input into the trained end-to-end model to obtain the result image marked with detection box and position information.

[0118] Qualitative evaluation refers to visualizing image detection results; quantitative evaluation refers to evaluating model performance through detection evaluation indicators; quantitative evaluation indicators include prediction accuracy AP3p25, detection accuracy Acc at recall rate = 0.9, average precision mAP50 when the intersection-over-union ratio is greater than 50, and average precision mAP50-95 when the intersection-over-union ratio is greater than 50 and less than 95.

[0119] The present invention proposes a new spectral perception network SA-YOLO to solve the key challenge that the YOLO model trained in visible light scenes cannot be directly applied to the infrared weak target detection task. First, the present invention designs the PSE_Former module, which significantly enhances the spectral feature expression ability of infrared weak targets through the triple structure of PAM, MSS and PGG, and solves the problem of infrared weak target feature submersion in traditional methods under low signal-to-noise ratio conditions. Secondly, the proposed SSPM adopts a progressive void rate and a dual-branch attention mechanism to achieve robust extraction of multi-scale features. Then, in response to the scale sensitivity of infrared targets, an innovative scale-adaptive DS-IoU loss is proposed, which reduces the positioning error of targets of different sizes through dynamic penalty weights and a composite normalization mechanism. Finally, the present invention constructs the ISUAV-D infrared UAV small target dataset. A large number of experimental verifications have shown that SA-YOLO has significant advantages in environments such as strong light, clouds, occlusion and interference, such as Figure 8 and Figure 9 In the future, the present invention plans to expand SA-YOLO to a wider range of infrared perception tasks such as multi-spectral collaborative detection and dynamic target tracking.

[0120] 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 aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for infrared dim small target detection based on cross-domain feature migration, characterized in that: The following steps are involved: 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; the self-built dataset is named ISUAV-D; S3, design the position-guided enhancement module PSE_Former: PSE_Former includes the position-aware module PAM, the multi-scale saliency fusion module MSS, and the position-guided filtering module PGG; PAM, MSS, and PGG are in a cascade relationship in PSE_Former; S4, design of spectral-spatial perception module SSPM: SSPM includes progressive dilation rate group strategy and local-global dual-path attention module LGAM; S5, design sensitivity stability loss function DS-IoU: The sensitivity stability loss function DS-IoU includes ordinary IoU and DS-IoU, and the two parts are composed of loss activation factors Decide, is a hyperparameter; S6, network reconstruction and model training: Use PSE_Former in step S3 and SSPM in step S4 to reconstruct a new network SA-YOLO on the basic YOLO network; use the training sets of the three datasets in steps S1 and S2, and the DS-IoU loss and cross entropy loss in step S5 for model training, and use the validation sets of the three datasets for model fine-tuning; S7, model testing and evaluation: After the fine-tuning in step S6 is completed, solidify the fine-tuned network parameters and save the final small target detection model; use the test sets in steps S1 and S2 to test the model effect from qualitative and quantitative perspectives. The real infrared image can be directly input into the trained end-to-end model to obtain the result image marked with detection box and position information.

2. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S1, the specific acquisition method and data structure of the NUST-SIRST dataset and the SIRST-5k dataset can be found in the existing description.

3. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S2, the data acquisition, data preprocessing and data partitioning of the ISUAV-D dataset are as follows: The ISUAV-D dataset is captured by a long-wave infrared camera. The target is a quadrotor drone, and the background is a daytime scene with buildings, sky, white clouds, pedestrians, and trees. It contains 16 image sequences. The ISUAV-D dataset belongs to the complex background interference dataset. The image format is PNG format and the image size is 640×512. The original ISUAV-D dataset contains a total of 20,000 images, including samples in which targets disappear for long periods of time. These samples were removed, and the final ISUAV-D dataset contains 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 high-order degradation model was used to perform two-stage simulated degradation on the images in the ISUAV-D dataset. No local degradation process was used. The degradation factors of the high-order degradation model should be selected to be small to prevent the target from being completely submerged in the background, and finally the degraded dataset was obtained.

4. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S3, the position perception module performs feature normalization, convolution mapping, position attention calculation, position encoding, scaled dot product calculation, convolution operation and activation function operation; the input of PAM is the output feature map of SPPF in YOLOv11 , First, after normalization and convolution mapping, three sets of matrix tensors Q, K, and V are generated for attention calculation. This process can be expressed as follows: ； in, Represents the position embedding matrix. The position encoding is embedded in the K vector as external explicit information to compensate for the lack of spatial perception information of the original Transformer. Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. 、 、 Respectively represent the query matrix weight, key matrix weight and value matrix weight; q, k, v represent the identification of the weight matrix function; The attention weight is calculated using the standard scaled dot product, introducing the regional spatial mask Controlling the focus range can be expressed using the formula: ； in, represents the output feature map of PAM, Represents the convolution operation, Softmax represents the normalized activation function, Represents the matrix corresponding element multiplication operation, represents the scaling factor, d represents the dimension of the key vector, and the square root is used to control the size of the dot product; Representation matrix The transpose operation; Finally, the attention outputs are integrated through the convolutional layer to generate enhanced feature maps.

5. The infrared small target detection method based on cross-domain feature migration 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 is composed of convolution layer 1, convolution layer 2 and convolution layer 3, 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; Output feature map of PAM At the same time, three convolution operations with different convolution kernel sizes are performed to extract fine-grained , medium particle size and coarse-grained Feature representation, Completely describes a three-dimensional feature graph structure, Indicates that each pixel value in the feature map 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 dynamic weighted subnetwork is designed to generate fusion weights , and , multi-scale saliency fusion features are obtained through weighted fusion .

6. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S3, the position-guided filtering module includes a background estimation network, a target gating layer 1, a target gating layer 2, and a depth-separable convolutional layer; First, multi-scale saliency fusion features Generate background response through background estimation and target gating network respectively and gating factors , and then get the intermediate features , this process can be expressed as: ； ； in, Represents a splicing operation, express Positional encoding, represents the gating network; In order to capture the contrast information of local areas, depthwise separable convolution is used to extract low-frequency responses. , and construct high-frequency residuals Finally, the enhanced salient features are further adjusted by introducing the contrast gating factor , this process can be expressed as: 。 7. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S4, the progressive hole rate group strategy establishes spectrum-spatial infrared band perception through frequency band adaptive hole rate; the hole rate group in SSPM The setting is progressive. The shallow features use a small void ratio to retain fine edges and local textures, and the deep features use a large void ratio to expand the receptive field and enhance the perception of the overall structure.

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

9. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S5, the design method of the sensitivity stability loss function DS-IoU is as follows: First, define the diagonal length of the target box As a benchmark, The formula is: ； in, and Represent the width and height of the target box respectively; Then, using the dynamic benchmark scale Constructing gradient compensation coefficients ; The formula is: ； in, and Respectively represent the width and height of the entire image; The scale penalty weight It can be expressed using the formula: ； Then the DS-IoU loss function can be defined as: ； in, represents the center offset penalty term, which inherits DIoU; Represents the size penalty term; it is expressed as follows: ； ； in, Represents the Euclidean distance between the center point of the predicted box and the center point of the real box; Represents the diagonal length of the minimum enclosing rectangle of the predicted box and the true box; and Represent the width and height of the real box respectively; and Represents the absolute difference between the width and height of the predicted box and the width and height of the real box respectively; DS-IoU loss function is obtained by Dynamically adjust the penalty scale. When the target size is smaller than the benchmark scale, Exponential decay makes the small target detection loss less sensitive to position offset; Normalization processing to eliminate the scale interference of the target size on the penalty term; set the IoU loss activation factor , to avoid over-optimization under high overlap.

10. The infrared small target detection method based on cross-domain feature migration according to claim 1, characterized in that: In S6, the basic YOLO network is a YOLOv11 network; the reconstruction method is to use a position-guided enhancement module PSE_Former and a spectral-spatial perception module SSPM to replace the C2PSA module and the C3k2 module in the YOLOv11 network to achieve infrared cross-domain feature migration; the classification loss is a cross-entropy loss.

Images