Infrared image small target detection method and system based on boundary perception and multi-stage aggregation network
By constructing a multi-stage aggregation module, a spatial recombination sampling module, and a spatial perception enhancement module, and combining boundary perception and a multi-stage aggregation network, the problems of feature loss and difficulty in localization in infrared small target detection are solved, and the accuracy of infrared small target detection is significantly improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-21
AI Technical Summary
Existing infrared small target detection methods struggle to effectively extract target features when faced with low contrast and complex background noise, leading to feature loss and difficulty in localization. In particular, traditional downsampling methods cannot retain fine-grained information and ignore the impact of spatial transformation on target detection.
A multi-stage aggregation module, a spatial re-sampling module, and a spatial perception enhancement module are constructed. Combining boundary perception and a multi-stage aggregation network, the multi-stage aggregation module fuses shallow details, intermediate layers, and deep semantic features, while spatial re-sampling preserves spatial and channel-level detail information. The spatial perception enhancement module performs multi-angle spatial perception and channel reweighting to establish a boundary perception and multi-stage aggregation network model, which is then optimized using training data.
It significantly improves the accuracy of infrared small target detection, effectively suppresses background noise, retains fine-grained information of small targets, enhances target boundary perception, and improves detection accuracy and robustness.
Smart Images
Figure CN121904530A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infrared image target detection technology, specifically to a method and system for detecting small targets in infrared images based on boundary perception and multi-stage aggregation networks. Background Technology
[0002] Detection of small infrared targets presents significant challenges, primarily due to low contrast, complex background noise, and blurred boundaries between the target and the background. Because these targets typically occupy small image areas and their signals are difficult to discern amidst background noise, traditional image processing methods struggle to effectively extract their detailed features, resulting in significant shortcomings in accuracy and robustness for existing detection methods. In recent years, with the rapid development of deep learning technology, many researchers have proposed various methods to improve the detection performance of small infrared targets. Existing research mainly focuses on enhancing the detection performance of small infrared targets by improving convolutional neural network structures and using multi-scale feature fusion. Despite these achievements, many challenges remain.
[0003] The low contrast of small infrared targets and the interference of background noise often cause existing convolutional neural networks to rely too heavily on global features when extracting target features, thus neglecting local details. While this global information extraction method helps the network understand the overall structure of the image, it fails to effectively capture the key features of small targets. Furthermore, information about small targets in infrared images is often significantly lost in high-level image features, especially when using traditional downsampling methods, where fine-grained target information cannot be effectively preserved. Moreover, existing networks often ignore the impact of spatial transformations on target detection during feature extraction, resulting in incomplete representation of target features. To enhance the network's sensitivity to small targets and suppress background interference, it is necessary to propose a structured scheme based on boundary awareness and multi-stage aggregation to improve the accuracy of infrared small target detection. Summary of the Invention
[0004] The technical problem to be solved by this invention is:
[0005] The purpose of this invention is to provide a method and system for detecting small targets in infrared images based on boundary awareness and multi-stage aggregation networks, in order to solve the problems of feature loss and difficulty in localization caused by low contrast and complex backgrounds in infrared small target detection.
[0006] The technical solution adopted by the present invention to solve the above problems is as follows:
[0007] (1) Construct a multi-stage aggregation module;
[0008] (2) Construct a spatial reassembly sampling module;
[0009] (3) Construct a spatial perception enhancement module;
[0010] (4) Establish a boundary-aware and multi-stage aggregation network model;
[0011] (5) Train the network using existing data;
[0012] (6) Use the trained model to detect infrared images.
[0013] Furthermore, the construction of the multi-stage aggregation module in step (1) above can be carried out according to the following steps:
[0014] The multi-stage aggregation module first projects the input features through a 1×1 convolution to twice the input channels 2c, and then divides them into left and right branches along the channel dimension: the left branch retains y0 to preserve shallow detail information; the right branch enters a feature extractor composed of n Bottleneck units connected in series to extract multi-stage semantic features, generating y0. (1) ,...,y (n) The multi-stage output of the right branch after each Bottleneck is concatenated with the left branch by channel to obtain a tensor with (2+n)c channels. This tensor is then compressed back to c channels by a 1×1 convolution to obtain the aggregated candidate features, denoted as y. c This feature contains multidimensional information from shallow, deep, and intermediate layers. The specific calculation is as follows:
[0015] (twenty one)
[0016] in, This represents a 1×1 convolution operation. This represents a splicing operation.
[0017] The module uses y0 as the value tensor basis, first unifies the channel dimension through several small convolutional kernels and linear transformations, and then extracts overlapping blocks by local k×k windows through an unfolding operation. Each block contains k 2 Each spatial location is then used to rearrange the shape of the value tensor according to the multi-head partitioning, denoted as V0.
[0018] The multi-stage aggregation module is further based on the intermediate layer y. m Weight the expanded value tensor V0. Using y m The attention tensor for each block is generated through a linear layer and reshaped to a size k per head. 2 ×k 2 The local interaction matrix is used to weight the V0 features with attention generated from the interaction matrix, resulting in an aggregated result after intermediate result emphasis:
[0019] (twenty two)
[0020] in, It is an interaction matrix generated by performing a linear transformation on the intermediate features. It is matrix multiplication.
[0021] Subsequently, using y c The second attention matrix A generated for the input c The above representation is then corrected again to obtain the final local redistribution result.
[0022] (twenty three)
[0023] in, This indicates the operation of reassembling the weighted blocks back into the feature map.
[0024] Finally, the fused output and the aggregated candidate features are summed using the residuals to obtain the final output:
[0025] (twenty four)
[0026] Furthermore, the spatial reassembly sampling in step (2) above can be performed as follows:
[0027] The spatial reconstruction sampling module first processes the input features using a 3×3 convolution while maintaining the spatial resolution. Then, it divides the data into four sub-maps using a 2×2 grid, each processed by a 1×1 convolution to obtain f1, f2, f3, and f4. The specific calculation is as follows:
[0028] (25)
[0029] (26)
[0030] (27)
[0031] (28)
[0032] in, This is the feature map before segmentation. This indicates that selections will begin from row 0 and column 0 with a step size of 2.
[0033] Next, these feature maps are concatenated along the channel dimension and fused using a 3×3 convolutional layer to obtain the final downsampled feature map. The specific calculation is as follows:
[0034] (29)
[0035] Furthermore, the spatial perception enhancement module in step (3) above can be performed according to the following steps:
[0036] The spatial perception enhancement module processes the spatial information of the input feature map x through two independent linear transformations. For horizontal spatial information, the input feature map undergoes independent linear projection to achieve feature recombination and transformation in the height direction, capturing dependencies in the height dimension. The formula is expressed as:
[0037] (30)
[0038] in This is a multilayer perceptron applied along the height dimension. For vertical spatial information, the feature map x undergoes another independent linear projection to achieve feature recombination and transformation in the width direction, as shown in the formula:
[0039] (31)
[0040] in It is a multilayer perceptron applied along the height dimension.
[0041] After obtaining the feature maps after horizontal and vertical transformations, the module concatenates these feature maps with the original input feature map x along the channel dimension to obtain the fused feature map. The formula is expressed as:
[0042] (32)
[0043] Furthermore, the module introduces a channel attention mechanism to adaptively adjust the importance of different channels. The channel attention vector is obtained through global average pooling, two convolutional layers, and sigmoid activation, specifically calculated as follows:
[0044] (33)
[0045] in, and These are convolutional layers used for dimensionality reduction and dimensionality enhancement, respectively. It is the Sigmoid activation function. Finally, the channel attention weights are multiplied element-wise with the fused feature map to obtain the weighted feature map:
[0046] (34)
[0047] in, This represents element-wise multiplication. The feature map after channel attention weighting is processed through a linear transformation layer to obtain the final feature map.
[0048] Further, in step (4), a boundary-aware and multi-stage aggregation network model is established. Convolutional layers and spatial reconstruction sampling modules are added to the shallow layers of the network backbone, followed by four multi-stage aggregation modules to extract features at different levels. A spatial reconstruction sampling module is added between the first two multi-stage aggregation modules. Convolutional layers are added between the remaining multi-stage aggregation modules. All convolutional layers consist of a standard 3×3 convolutional kernel, a batch normalization layer, and a Silu activation function. Each multi-stage aggregation module outputs feature maps with 128, 256, 512, and 1024 channels, respectively. The last feature map is processed by fast spatial pyramid pooling. The extracted feature maps at different levels are used to construct a feature pyramid, and the feature pyramid is fused from top to bottom. The fused result is enhanced by the spatial awareness enhancement module and then input into the YOLO detection head to obtain the detection box.
[0049] Furthermore, the training of the network using existing data in step (5) above can be performed as follows:
[0050] The constructed network is trained using an infrared small target detection dataset. During training, the network parameters are updated in reverse using a loss function until the optimal parameters are obtained. The network's loss function consists of a weighted average of classification and regression losses. The classification loss uses binary cross-entropy. The calculation formula is as follows:
[0051] (35)
[0052] in, It is a binary tag. It outputs the probability of belonging to the label. This indicates the number of groups of objects predicted by the model.
[0053] The regression loss is calculated using a combination of GIoU and distributed focal loss. GIoU penalizes the bias of the predicted bounding box by minimizing the bounding box area, helping the model to gradually correct the position and size of the predicted box. The formula for calculating GIoU is as follows:
[0054] (36)
[0055] (37)
[0056] Where A is the predicted bounding box, B is the true bounding box, and C is the smallest rectangle that contains A and B.
[0057] The formula for calculating distributed focus loss is as follows:
[0058] (38)
[0059] in, and These are the predicted values output by the network, and the nearest neighbor predicted values. , , The values are the actual value of the label, the label integral value, and the integral value of the neighboring labels.
[0060] The final loss function is calculated as follows:
[0061] (39)
[0062] The infrared small target dataset was divided into training, validation, and test sets in a 7:1:2 ratio. The SGD optimizer was chosen as the network optimizer. The initial learning rate was 0.01, which decayed exponentially in each iteration. The model was trained for a total of 200 generations, with each iteration...
[0063] Furthermore, step (7) above can be performed as follows:
[0064] The infrared image to be detected is input into the trained network to obtain the detection result. This process can be represented as:
[0065] (40)
[0066] in, This represents the constructed boundary-aware and multi-stage aggregation network. and These represent the input infrared image and the output detection box result, respectively.
[0067] The present invention has the following beneficial technical effects:
[0068] The present invention includes the following steps: (1) constructing a multi-stage aggregation module; (2) constructing a spatial reconstructed sampling module; (3) constructing a spatial perception enhancement module; (4) establishing a boundary perception and multi-stage aggregation network model; (5) training the network using existing data; and (6) using the trained model to detect infrared images. The present invention can effectively improve the detection accuracy of small infrared targets by strengthening boundary perception and performing multi-stage aggregation when faced with the problems of feature loss and difficulty in localization caused by low contrast and complex backgrounds of small infrared targets.
[0069] This invention presents a method for small target detection in infrared images based on boundary awareness and a multi-stage aggregation network. By organically aggregating shallow details, intermediate layers, and deep semantics, this invention enhances target boundary awareness and suppresses background noise. Through spatial resampling, it retains more spatial and channel-level detail information during the shallow downsampling stage, reducing feature loss for small targets at the source. Furthermore, by modeling lateral and longitudinal spatial information separately and combining channel attention, this invention achieves multi-angle spatial awareness and channel reweighting, improving the ability to discriminate small targets after multi-scale fusion. Experiments show that, compared with existing infrared small target detection methods, this invention effectively suppresses background noise, retains fine-grained information about small targets, and significantly improves the detection accuracy of infrared small targets. Attached Figure Description
[0070] Figure 1 This is a flowchart of the method of the present invention.
[0071] Figure 2 This is a schematic diagram of the overall structure of the network and the multi-stage aggregation module proposed in this invention (a schematic diagram of the overall structure of the infrared image small target detection method based on boundary sensing and multi-stage aggregation network proposed in this invention).
[0072] Figure 3 This is a schematic diagram of the spatial reconstruction sampling module.
[0073] Figure 4 This is a schematic diagram of the spatial perception enhancement module structure;
[0074] Figure 5 This is a comparison chart of the detection results of the method of this invention with other currently advanced methods; Detailed Implementation
[0075] The present invention will now be described in detail with reference to the accompanying drawings and examples.
[0076] See the flowchart of the method of this invention. Figure 1 The specific implementation steps are as follows:
[0077] (1) Construct a multi-stage aggregation module;
[0078] (2) Construct a spatial reassembly sampling module;
[0079] (3) Construct a spatial perception enhancement module;
[0080] (4) Establish a boundary-aware and multi-stage aggregation network model;
[0081] (5) Train the network using existing data;
[0082] (6) Use the trained model to detect infrared images.
[0083] The above step (1) is performed as follows:
[0084] The multi-stage aggregation module first projects the input features through a 1×1 convolution to twice the input channels 2c, and then divides them into left and right branches along the channel dimension: the left branch retains y0 to preserve shallow detail information; the right branch enters a feature extractor composed of n Bottleneck units connected in series to extract multi-stage semantic features, generating y0. (1) ,...,y (n) The multi-stage output of the right branch after each Bottleneck is concatenated with the left branch by channel to obtain a tensor with (2+n)c channels. This tensor is then compressed back to c channels by a 1×1 convolution to obtain the aggregated candidate features, denoted as y. c This feature contains multidimensional information from shallow, deep, and intermediate layers. The specific calculation is as follows:
[0085] (41)
[0086] in, This represents a 1×1 convolution operation. This represents a splicing operation.
[0087] The module uses y0 as the value tensor basis, first unifies the channel dimension through several small convolutional kernels and linear transformations, and then extracts overlapping blocks by local k×k windows through an unfolding operation. Each block contains k 2 Each spatial location is then used to rearrange the shape of the value tensor according to the multi-head partitioning, denoted as V0.
[0088] The multi-stage aggregation module is further based on the intermediate layer y. m Weight the expanded value tensor V0. Using y m The attention tensor for each block is generated through a linear layer and reshaped to a size k per head. 2 ×k 2 The local interaction matrix is used to weight the V0 features with attention generated from the interaction matrix, resulting in an aggregated result after intermediate result emphasis:
[0089] (42)
[0090] in, It is an interaction matrix generated by performing a linear transformation on the intermediate features. It is matrix multiplication.
[0091] Subsequently, using y c The second attention matrix A generated for the input c The above representation is then corrected again to obtain the final local redistribution result.
[0092] (43)
[0093] in, This indicates the operation of reassembling the weighted blocks back into the feature map.
[0094] Finally, the fused output and the aggregated candidate features are summed using the residuals to obtain the final output:
[0095] (44)
[0096] Step (2) above shall be performed as follows:
[0097] The spatial reconstruction sampling module first processes the input features using a 3×3 convolution while maintaining the spatial resolution. Then, it divides the data into four sub-maps using a 2×2 grid, each processed by a 1×1 convolution to obtain f1, f2, f3, and f4. The specific calculation is as follows:
[0098] (45)
[0099] (46)
[0100] (47)
[0101] (48)
[0102] in, This is the feature map before segmentation. This indicates that selections will begin from row 0 and column 0 with a step size of 2.
[0103] Next, these feature maps are concatenated along the channel dimension and fused using a 3×3 convolutional layer to obtain the final downsampled feature map. The specific calculation is as follows:
[0104] (49)
[0105] Step (3) above shall be performed as follows:
[0106] The spatial perception enhancement module processes the spatial information of the input feature map x through two independent linear transformations. For horizontal spatial information, the input feature map undergoes independent linear projection to achieve feature recombination and transformation in the height direction, capturing dependencies in the height dimension. The formula is expressed as:
[0107] (50)
[0108] in This is a multilayer perceptron applied along the height dimension. For vertical spatial information, the feature map x undergoes another independent linear projection to achieve feature recombination and transformation in the width direction, as shown in the formula:
[0109] (51)
[0110] in It is a multilayer perceptron applied along the height dimension.
[0111] After obtaining the feature maps after horizontal and vertical transformations, the module concatenates these feature maps with the original input feature map x along the channel dimension to obtain the fused feature map. The formula is expressed as:
[0112] (52)
[0113] Furthermore, the module introduces a channel attention mechanism to adaptively adjust the importance of different channels. The channel attention vector is obtained through global average pooling, two convolutional layers, and sigmoid activation, specifically calculated as follows:
[0114] (53)
[0115] in, and These are convolutional layers used for dimensionality reduction and dimensionality enhancement, respectively. It is the Sigmoid activation function. Finally, the channel attention weights are multiplied element-wise with the fused feature map to obtain the weighted feature map:
[0116] (54)
[0117] in, This represents element-wise multiplication. The feature map after channel attention weighting is processed through a linear transformation layer to obtain the final feature map.
[0118] The above step (4) shall be performed as follows:
[0119] The boundary-aware and multi-stage aggregation network adds convolutional layers and a spatial reconstruction sampling module to the shallow layers of the network backbone, followed by four multi-stage aggregation modules to extract features at different levels. A spatial reconstruction sampling module is added between the first two multi-stage aggregation modules. Convolutional layers are added between the remaining multi-stage aggregation modules. All convolutional layers consist of a standard 3×3 convolutional kernel, a batch normalization layer, and a Silu activation function. Each multi-stage aggregation module outputs feature maps with 128, 256, 512, and 1024 channels, respectively. The last feature map undergoes fast spatial pyramid pooling. The extracted feature maps at different levels are constructed into a feature pyramid, which is then fused from top to bottom. The fused result is enhanced by the spatial awareness enhancement module and input into the YOLO detection head to obtain detection boxes.
[0120] Step (5) above shall be performed as follows:
[0121] The constructed network is trained using an infrared small target detection dataset. During training, the network parameters are updated in reverse using a loss function until the optimal parameters are obtained. The network's loss function consists of a weighted average of classification and regression losses. The classification loss uses binary cross-entropy. The calculation formula is as follows:
[0122] (55)
[0123] in, It is a binary tag. It outputs the probability of belonging to the label. This indicates the number of groups of objects predicted by the model.
[0124] The regression loss is calculated using a combination of GIoU and distributed focal loss. GIoU penalizes the bias of the predicted bounding box by minimizing the bounding box area, helping the model to gradually correct the position and size of the predicted box. The formula for calculating GIoU is as follows:
[0125] (56)
[0126] (57)
[0127] Where A is the predicted bounding box, B is the true bounding box, and C is the smallest rectangle that contains A and B.
[0128] The formula for calculating distributed focus loss is as follows:
[0129] (58)
[0130] in, and These are the predicted values output by the network, and the nearest neighbor predicted values. , , The values are the actual value of the label, the label integral value, and the integral value of the neighboring labels.
[0131] The final loss function is calculated as follows:
[0132] (59)
[0133] The infrared small target dataset was divided into training, validation, and test sets in a 7:1:2 ratio. The SGD optimizer was chosen as the network optimizer. The initial learning rate was 0.01, which decayed exponentially in each iteration. The model was trained for a total of 200 generations, with each iteration...
[0134] The above step (6) shall be performed as follows:
[0135] The infrared image to be detected is input into the trained network to obtain the detection result. This process can be represented as:
[0136] (60)
[0137] in, This represents the constructed boundary-aware and multi-stage aggregation network. and These represent the input infrared image and the output detection box result, respectively.
[0138] To quantitatively evaluate the performance of the method proposed in this invention, the experiments used precision, recall, accuracy (AP) for each target class, and mean accuracy (mAP) for evaluation. Precision is defined as follows:
[0139] (61)
[0140] Where TP is the number of positive classes predicted as positive, and FP is the number of negative classes predicted as positive. Recall is defined as follows:
[0141] (62)
[0142] Where FN is the number of positive classes predicted as negative classes. Precision AP is obtained by calculating the area integral under the precision-recall curve. Mean precision mAP is the average precision for each class of targets. The numerical subscripts of mean precision mAP represent different intersection-union (IU) thresholds.
[0143] The method proposed in this invention is compared with RT-DETR, YOLOv10, YOLOv11, YOLOv12 and ABRNet respectively. Figure 5The results visualization of the proposed method and other state-of-the-art methods is presented. Targets that were missed are circled in red, and major false positive areas are marked in yellow. Under low contrast and complex background interference, other comparative methods missed dimly lit targets. Through boundary awareness and multi-stage aggregation, the proposed method can more accurately distinguish between targets and background. The proposed method achieved the best detection results in various scenarios.
[0144] Table 1. Quantitative comparison results of different methods in infrared small target detection on the test dataset.
[0145]
[0146] Table 1 lists the quantitative comparison results of the method of this invention with other current advanced detection methods, including precision, recall, accuracy (AP) for each target class, and mean accuracy (mAP). In terms of the most important mAP, the method of this invention outperforms all other methods. The quantitative comparison results show that the method proposed in this invention exhibits the best performance and can significantly improve the detection accuracy of small infrared targets.
Claims
1. A method for detecting small targets in infrared images based on boundary awareness and multi-stage aggregation networks, characterized in that, The method improves the accuracy of infrared small target detection by enhancing boundary perception of small targets to improve their separability in complex backgrounds and by performing multi-stage aggregation during feature extraction to suppress complex background noise. The method includes the following steps: (1) Construct a multi-stage aggregation module; (2) Construct a spatial reorganization sampling module; (3) Construct a spatial perception enhancement module; (4) Establish a boundary-aware and multi-stage aggregation network model; (5) Train the network using existing data; (6) Use the trained model to detect infrared images.
2. The method as described in claim 1, characterized in that, Step (1) is performed as follows: The multi-stage aggregation module first projects the input features through a 1×1 convolution to twice the input channels 2c, and then divides them into left and right branches along the channel dimension: the left branch retains y0 to preserve shallow detail information; the right branch enters a feature extractor composed of n Bottleneck units connected in series to extract multi-stage semantic features, generating y0. (1) ,...,y (n) The multi-stage output of the right branch after each Bottleneck is concatenated with the left branch by channel to obtain a tensor with (2+n)c channels. This tensor is then compressed back to c channels by a 1×1 convolution to obtain the aggregated candidate features, denoted as y. c This feature contains multidimensional information from shallow, deep, and intermediate layers, and its specific calculation is as follows: （1） in, This represents a 1×1 convolution operation. This represents a splicing operation; The module uses y0 as the value tensor basis, first unifies the channel dimension through several small convolutional kernels and linear transformations, and then extracts overlapping blocks by local k×k windows through an unfolding operation. Each block contains k 2 A spatial location; then rearrange the shape of the value tensor according to the multi-head partition, denoted as V0. The multi-stage aggregation module is based on the intermediate layer y m Weight the expanded value tensor V0 and use y m The attention tensor for each block is generated through a linear layer and reshaped to a size k per head. 2 ×k 2 The local interaction matrix is used to weight the V0 features with attention generated from the interaction matrix, resulting in an aggregated result after intermediate result emphasis: （2） in, It is an interaction matrix generated by performing a linear transformation on the intermediate features. It is matrix multiplication; Subsequently, using y c The second attention matrix A generated for the input c The above representation is then corrected again to obtain the final local redistribution result; （3） in, This indicates the operation of reassembling the weighted blocks back into the feature map; Finally, the fused output and the aggregated candidate features are summed using the residuals to obtain the final output: （4）。 3. The method as described in claim 1, characterized in that, Step (2) is performed as follows: The spatial reconstruction sampling module first processes the input features using a 3×3 convolution while maintaining the spatial resolution. Then, it divides the input features into four sub-maps using a 2×2 grid, each processed by a 1×1 convolution to obtain f1, f2, f3, and f4. The specific calculation is as follows: （5） （6） （7） （8） in, This is the feature map before segmentation. This indicates that selections will begin from row 0 and column 0 with a step size of 2. Next, these feature maps are concatenated along the channel dimension and fused using a 3×3 convolutional layer to obtain the final downsampled feature map. The specific calculation is as follows: （9）。 4. The method as described in claim 1, characterized in that, Step (3) is performed as follows: The spatial perception enhancement module processes the spatial information of the input feature map x through two independent linear transformations. For horizontal spatial information, the input feature map undergoes independent linear projection to achieve feature recombination and transformation in the height direction, capturing the dependencies in the height dimension. The formula is expressed as: （10） in This is a multilayer perceptron applied along the height dimension. For vertical spatial information, the feature map x undergoes another independent linear projection to achieve feature recombination and transformation in the width direction, as shown in the formula: （11） in It is a multilayer perceptron applied along the height dimension; After obtaining the feature maps after horizontal and vertical transformations, the module concatenates these feature maps with the original input feature map x along the channel dimension to obtain the fused feature map, which is expressed by the formula: （12） Furthermore, the module introduces a channel attention mechanism to adaptively adjust the importance of different channels; the channel attention vector is obtained through global average pooling, two layers of convolution, and sigmoid activation, and the specific calculation is as follows: （13） in, and These are convolutional layers used for dimensionality reduction and dimensionality enhancement, respectively. It uses the Sigmoid activation function; finally, the channel attention weights are multiplied element-wise with the fused feature map to obtain the weighted feature map: （14） in, This indicates element-wise multiplication. The feature map after channel attention weighting is processed by a linear transformation layer to obtain the final feature map.
5. The method as described in claim 1, characterized in that, Step (4) is performed as follows: The boundary-aware and multi-stage aggregation network adds convolutional layers and a spatial reconstruction sampling module to the shallow layers of the network backbone, followed by four multi-stage aggregation modules to extract features at different levels. A spatial reconstruction sampling module is added between the first two multi-stage aggregation modules; convolutional layers are added between the remaining multi-stage aggregation modules. All convolutional layers consist of a 3×3 standard convolutional kernel, a batch normalization layer, and a Silu activation function. Each multi-stage aggregation module outputs feature maps with 128, 256, 512, and 1024 channels, respectively. The last feature map undergoes fast spatial pyramid pooling to construct a feature pyramid from the extracted feature maps at different levels. The feature pyramids are then fused from top to bottom. The fused result is enhanced by the spatial awareness enhancement module and input into the YOLO detection head to obtain the detection box.
6. The method as described in claim 5, characterized in that, Step (5) is performed as follows: The constructed network is trained using an infrared small target detection dataset. During training, the network parameters are updated in reverse using a loss function until the optimal network parameters are obtained. The network's loss function consists of a weighted average of classification loss and regression loss. The classification loss uses binary cross-entropy, and the calculation formula is as follows: （15） in, It is a binary tag. It outputs the probability of belonging to the label. This indicates the number of groups of objects predicted by the model; The regression loss is calculated using a combination of GIoU and distributed focal loss. GIoU penalizes the bias of the predicted box by minimizing the bounding box area, helping the model to gradually correct the position and size of the predicted box. The formula for calculating GIoU is as follows: （16） （17） Where A is the predicted bounding box, B is the true bounding box, and C is the smallest rectangle that contains A and B. The formula for calculating distributed focus loss is as follows: （18） in, and The network outputs predicted values, and the nearest neighbor predicted values. , , The values are the actual value of the label, the label integral value, and the integral value of the neighboring labels. The final loss function is calculated as follows: （19） The infrared small target dataset was divided into a training set, a validation set, and a test set in a ratio of 7:1:2; the SGD optimizer was selected as the network optimizer; the initial learning rate during training was 0.01, which decayed exponentially in each iteration; the model was trained for a total of 200 generations, with each iteration...
7. The method as described in claim 1, characterized in that, Step (6) is performed as follows: The infrared image to be detected is input into the trained network to obtain the detection result. This process can be represented as: （20） in, This represents the constructed boundary-aware and multi-stage aggregation network. and These represent the input infrared image and the output detection box result, respectively.
8. A small target detection system based on boundary sensing and multi-stage aggregation network in infrared imagery, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-8 above, and executes the steps in the infrared image small target detection method based on boundary perception and multi-stage aggregation network described above when running.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the infrared image small target detection method based on boundary sensing and multi-stage aggregation network as described in any one of claims 1-8.