An infrared dim small target cooperative enhancement detection method based on improved YOLO11n

Abstract

The application discloses an infrared dim small target cooperative enhancement detection method based on an improved YOLO11n, and comprises the following steps: 1, constructing an infrared dim small target data set, and dividing a training set, a test set and a verification set; 2, building an infrared dim small target cooperative enhancement detection network ISTCE-Net; 3, training the detection network ISTCE-Net by using the training set data; 4, evaluating the performance of the detection network ISTCE-Net by using the verification set and optimizing parameters; and 5, inputting a to-be-detected image into the trained detection network ISTCE-Net to obtain an infrared dim small target detection result. The application can solve the problems in the prior art that infrared dim small target features are weak, background thermal noise interference is strong, multi-scale feature transmission is prone to losing key pixels, and classification confidence is difficult to reflect positioning quality.

Description

Technical Field

[0001] This invention belongs to the field of infrared image intelligent analysis and target detection technology, and relates to a method for collaborative enhancement detection of weak infrared targets based on improved YOLO11n. Background Technology

[0002] Infrared imaging forms images by sensing the difference in thermal radiation between a target and its surrounding environment. It is less affected by lighting conditions and can operate at night or in low-visibility environments, making it widely used in road monitoring, drone inspections, remote sensing, maritime search and rescue, and public safety. However, due to factors such as imaging distance, sensor resolution, and complex environmental interference, the number of pixels occupied by the target in an infrared image is usually small. Furthermore, it suffers from low contrast, blurred edges, insufficient texture information, and susceptibility to being overwhelmed by background thermal noise, making the detection of small infrared targets still quite challenging.

[0003] In recent years, with the development of deep learning technology, object detection methods based on convolutional neural networks have gradually replaced traditional methods that rely on manually designed features. Existing object detection methods mainly include two-stage and single-stage methods. Two-stage methods first generate candidate regions, then classify and regress their locations, typically achieving high detection accuracy but incurring significant computational overhead. Single-stage methods directly predict the object category and location on multi-scale feature maps, offering better detection efficiency and making them more suitable for real-time detection scenarios. As a representative algorithm in the field of single-stage detection, the YOLO11n single-stage object detection network mainly consists of a backbone feature extraction network, a multi-scale feature fusion network, and a detection head. It completes object category prediction and bounding box location regression through a single forward propagation, achieving a good balance between model size, inference speed, and detection accuracy.

[0004] However, the YOLO11n single-stage target detection network is primarily designed for general visible light images. When directly applied to infrared weak target detection, it still has certain limitations. First, existing networks mainly rely on local convolution operations, making it difficult to simultaneously utilize large-scale background information and local edge details of the target, and are easily affected by large-area thermal noise. Second, during multi-scale feature fusion, traditional downsampling operations may cause the loss of key pixels of weak targets, and conventional upsampling and simple feature stitching are also difficult to effectively suppress background thermal noise. Third, existing detection heads typically handle target classification and bounding box regression separately, lacking a dynamic evaluation mechanism for localization quality for infrared blurred boundaries, which may result in high classification confidence but insufficient localization accuracy.

[0005] In summary, existing single-stage target detection networks still have technical problems in infrared weak target detection, such as difficulty in taking into account both global background information and local edge details, easy loss of key pixels during multi-scale feature transfer, and mismatch between classification confidence and localization quality. Summary of the Invention

[0006] The purpose of this invention is to provide a collaborative enhancement detection method for infrared weak targets based on improved YOLO11n, which solves the problems in the prior art such as weak infrared target features, strong background thermal noise interference, easy loss of key pixels in multi-scale feature transmission, and difficulty in reflecting positioning quality by classification confidence.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A collaborative enhancement detection method for infrared weak targets based on improved YOLO11n includes the following steps: Step 1: Construct an infrared weak target dataset and divide it into training, testing, and validation sets; Step 2: Construct the infrared weak target collaborative enhancement detection network ISTCE-Net; The detection network ISTCE-Net is based on YOLO11n and includes a backbone feature extraction network, a space-frequency collaborative multi-scale feature aggregation network SFCA-Neck, and a positioning quality dynamic calibration detection head DLQC-Head. The backbone feature extraction network is configured with a deep global-local context feature enhancement module (GLCFE-Block). The GLCFE-Block adopts a structure that combines branch connections, cross-stage partial connections, and residual connections, including a cross-stage partial connection context enhancement unit (CSPCE) and a global-local dual-branch attention unit (GLDBA). The space-frequency collaborative multi-scale feature aggregation network SFCA-Neck includes a space-frequency three-branch dynamic calibration module SFDC-Block and a channel shift hybrid feature reconstruction upsampling module SCMU-Block. The space-frequency three-branch dynamic calibration module SFDC-Block adopts a feature partitioning and fusion structure with residual connections. The positioning quality dynamic calibration detection head DLQC-Head includes a classification branch, a position regression branch, and a regression distribution statistical quality assessment unit RDQA. Step 3: Train the detection network ISTCE-Net using the training set data from Step 1; Step 3.1: Input the infrared image to be detected into the backbone feature extraction network, and extract feature maps of different resolutions layer by layer. The output of the backbone feature extraction network includes at least shallow feature maps. Mid-layer feature map and deep feature map ,in Spatial resolution lower than and Furthermore, it contains richer target semantics and background context information; Step 3.2: The deep feature map output in Step 3.1 As input, global-local context enhancement is performed through the Global-Local Context Feature Enhancement (GLCFE-Block) module; Step 3.3: The space-frequency collaborative multi-scale feature aggregation network SFCA-Neck receives the shallow feature map output in step 3.1. Mid-layer feature map and the deep feature map enhanced by the GLCFE-Block output in step 3.2. The system performs bidirectional cross-layer feature fusion: in the top-down path, high-level features are upsampled to the same spatial resolution as the next layer features via SCMU-Block, then concatenated with the features of that layer along the channel dimension and calibrated via SFDC-Block; in the bottom-up path, the fused low-level features are downsampled to the same spatial resolution as the next layer features via spatial information rearrangement convolution, then concatenated with the features of that layer and calibrated via SFDC-Block, thereby outputting a multi-scale feature map that combines high-resolution spatial details and high-level semantic information. Step 3.4: Input the fused multi-scale feature map into the localization quality dynamic calibration detection head DLQC-Head. The localization quality dynamic calibration detection head DLQC-Head generates a localization quality calibration value based on the statistical characteristics of the bounding box regression probability distribution, and uses the calibration value to adjust the classification confidence, outputting the target category, bounding box position and corresponding confidence. Step 4: Use the validation set to evaluate the performance of the detection network ISTCE-Net and optimize the parameters; Step 5: Input the image to be detected into the trained detection network ISTCE-Net to obtain the infrared weak target detection results.

[0008] Furthermore, step 3.2 includes the following process: The deep feature map output in step 3.1 This is used as the input feature map of GLCFE-Block. ,in , , and These represent the height, width, and number of channels of the feature map, respectively; within CSPCE, first... Divided into bypass features along the channel dimension and context-enhanced features ; Reserved as a cross-stage bypass. Input GLDBA and further divide it into global features along the channel dimension. and local features The data are fed into the global branch and the local branch for parallel processing, respectively. Global branch receives global features Attention computation is constructed using local windows as units; specifically, query features are first obtained through linear mapping. : in This represents the linear mapping matrix of the query features; Again Perform average pooling and obtain key features via linear mapping. Sum value characteristics : in This indicates the average pooling operation performed within the attention window. and These represent linear mapping matrices for key features and value features, respectively; Finally, according to The relevance is used to calculate attention weights, and the relevance is used to calculate attention weights. We perform weighted summation to obtain the global branch output features. : in The channel dimension or scaling factor representing the key feature; After linear mapping, the local features are input into two parallel depthwise convolutional branches to obtain local response features. and Subsequently, key high-frequency features were screened through element-wise multiplication, nonlinear activation, and gating operations. in, This indicates element-wise multiplication. Represents a non-linear activation function. Represents the hyperbolic tangent function. Represents a scaled-down view of local features. Indicates the output characteristics of local branches; Finally, output the global branch. With local branch output The output features are obtained by concatenating along the channel dimension and then using linear mapping and residual connections. In CSPCE, the aforementioned As a context-enhanced branch output, it is compared with bypass features. After being stitched along the channel dimension, Convolution performs channel compression and feature fusion to obtain the CSPCE output features. Then It is fed into subsequent layers of GLCFE-Block to achieve synergistic enhancement of global background information and local edge detail information.

[0009] Furthermore, step 3.3 includes the following process: In the SFDC-Block spatial frequency three-branch dynamic calibration module, the input features are first processed by convolution and then divided into residual features and features to be calibrated along the channel dimension. The features to be calibrated are input into the global frequency branch, the large kernel spatial branch, and the local detail branch, respectively. The global frequency branch extracts the global energy distribution of the image through fast Fourier transform and inverse fast Fourier transform. The large kernel spatial branch uses 31×1, 1×31, and 31×31 depth convolutions to extract long-distance spatial dependencies in different directions. The local detail branch uses 1×1 depth convolutions to preserve pixel-level local information. in, This represents the output characteristics of the global frequency branch. This indicates the output characteristics of the large core spatial branch. This represents the output features of local detail branches. Representing residual characteristics, This represents the calibrated feature map output by SFDC-Block; In the channel shift hybrid feature reconstruction upsampling module SCMU-Block, the input features are first upsampled and depth-separable convolutional processed, and then divided into multiple feature groups along the channel dimension. After spatial displacement and channel blending operations, different feature groups are mapped to the output space using pointwise convolution. During the bottom-up feature transfer process, the SFCA-Neck employs a spatial information rearrangement downsampling operator to transfer spatial information from shallow feature maps to the channel dimension, thereby reducing the loss of key pixels of weak targets during the downsampling process.

[0010] Furthermore, the spatial information rearrangement downsampling operator is SPD-Conv.

[0011] Furthermore, step 3.4 includes the following process: The classification branch uses two concatenated depthwise separable convolutions and a nonlinear activation function to extract classification features; let the first... The input feature map for each detection scale is The output of two depthwise separable convolutions is ,in This represents depthwise separable convolution. Represents a non-linear activation function; then through Convolution yields the initial classification score. , its first The value of the channel corresponds to the _th On the first detection scale Initial classification score for the target class; The location regression branch does not directly output a single coordinate value, but instead outputs the offset of each position of the bounding box in a discrete interval. The probability distribution within; Let the output feature of the location regression branch be Then, the regression probability distribution is obtained through normalization: in, This represents the discrete regression feature of the position regression branch output. Indicates to conduct The regression probability distribution obtained after normalization, where reg_max represents the maximum index of the discrete regression interval; The regression distribution statistical quality assessment unit (RDQA) does not perform complex entropy calculations, but instead extracts the top k maximum values ​​and their mean from the regression probability distribution: Subsequently, the statistical feature vector Input to the multilayer sensor to generate positioning quality calibration values: Positioning quality calibration value Compared with the initial classification score logit Add them together to obtain the calibrated classification logit: in, This represents the maximum number of responses selected from the regression probability distribution, which can be set according to the requirements of detection accuracy and computational load. express Center front One maximum probability value; express The mean; Indicates by and The resulting location quality statistical feature vector is obtained by splicing. This represents the positioning quality calibration value generated by the MLP; This represents the classification logit after positioning quality calibration. implement The function yields the final classification confidence score used for output.

[0012] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a multi-layered feature map sequence from shallow to deep using a backbone feature extraction network, and introduces a global-local contextual feature enhancement mechanism in the deeper layers of the backbone feature extraction network, effectively improving the network's feature representation capability for infrared weak targets. Specifically, through the cooperation of cross-stage contextual enhancement units (CSPCE) and cross-stage bypasses, the original feature information is preserved while performing deep feature transformation, reducing the loss of details of weak targets. In CSPCE, a global-local dual-branch attention unit (GLDBA) is used to further process the input features: the global branch first performs a linear mapping on the features, and then aggregates key and numerical features within a local window using average pooling. It then calculates global context features reflecting a wide range of background relationships based on the attention between the query, key, and value. The local branch, after performing a linear mapping on the features, uses two parallel depthwise separable convolutional branches to extract local spatial features. The outputs of the two branches are multiplied element-wise to form a non-linear gating layer, which is then passed through a fully connected layer, a Swish activation function, and a Tanh activation function to generate a local detail scaling map, enhancing the high-frequency edge response of weak targets. The outputs of the global and local branches are concatenated along the channel dimension, and then feature fusion and channel restoration are completed through a fully connected layer. A residual connection is then established with the input features to obtain the global-local enhanced features. Finally, the global-local enhanced features are concatenated and fused with the cross-stage bypass features to output an enhanced deep feature map, laying a solid foundation for subsequent detection tasks.

[0013] This invention achieves the collaborative fusion of spatial, frequency, and multi-scale features through SFCA-Neck, effectively solving the problem of key pixels of weak targets being easily lost during feature transfer. The SFDC-Block dynamic calibration module with three branches of spatial and frequency frequencies suppresses background thermal noise, and the SCMU-Block upsampling module for channel-shifted hybrid feature reconstruction and the Space-to-Depth Convolution (SPD-Conv) downsampling operation transfer feature information at different scales. Specifically, the SPD-Conv segment the input feature map requiring a 2x downsampling into four sub-feature maps based on the parity of the row and column indices: even row even column, odd row even column, even row odd column, and odd row odd column. These four sub-feature maps are then concatenated along the channel dimension, allowing pixel information in each sub-feature map to be transferred losslessly from the spatial dimension to the channel dimension. Subsequently, convolutional layers are used for channel fusion and feature extraction, thereby reducing the spatial resolution of the feature map while retaining fine-grained pixel information related to infrared weak targets in the shallow features.

[0014] This invention establishes a dynamic correlation between classification confidence and positioning quality through DLQC-Head, significantly improving the reliability of detection results. The detection head can generate a positioning quality calibration value based on the statistical characteristics of the bounding box regression probability distribution, and use this calibration value to adjust the classification confidence. It outputs the target category, bounding box position, and corresponding confidence, ensuring that the final detection score simultaneously reflects both the category judgment result and positioning reliability. This reduces the impact of high-confidence false detections and low-quality positioning results, ensuring accurate category and precise position detection results even in complex infrared scenes, significantly reducing false alarm and false negative rates.

[0015] This invention enhances the ability to distinguish between weak infrared targets and background thermal noise by simultaneously strengthening global background context information and local edge detail information through GLCFE-Block. It utilizes SFCA-Neck to achieve the collaborative fusion of spatial domain, frequency domain, and multi-scale features, alleviating the problem of key pixels of weak targets being lost during upsampling and downsampling. At the same time, it uses DLQC-Head to dynamically calibrate the classification confidence based on the bounding box regression distribution, allowing the detection score to comprehensively reflect the category determination result and the reliability of target localization, effectively suppressing high-confidence false detections, reducing low-quality localization output, and improving the overall performance of infrared weak target detection. Attached Figure Description

[0016] Figure 1 This is a diagram of the overall architecture of the infrared weak target collaborative enhancement detection network ISTCE-Net of the present invention; Figure 2 A schematic diagram of the Global-Local Context Feature Enhancement (GLCFE-Block) module and its internal structure in the deep layers of the backbone network is shown below. Figure 2 (a) is GLCFE-Block. Figure 2 (b) is the Cross-Stage Context Enhancement Unit (CSPCE). Figure 2 (c) is the global-local dual-branch attention unit GLDBA; Figure 3 This is a schematic diagram of the SFDC-Block three-branch space-frequency dynamic calibration module. Figure 4 A schematic diagram of the SCMU-Block upsampling module for channel shift hybrid feature reconstruction; Figure 5 A schematic diagram of the DLQC-Head, a dynamic calibration head for positioning quality. Figure 6 A schematic diagram showing the detection results of the ISTCE-Net network on VEDAI infrared aerial imagery test images. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The infrared weak target cooperative enhancement detection method based on the improved YOLO11n described in this embodiment includes the following steps: Step 1: Construct an infrared weak target dataset, dividing it into training, testing, and validation sets. Step 2: Construct the Infrared Weak Target Cooperative Enhancement Detection Network ISTCE-Net like Figure 1 As shown, the detection network ISTCE-Net is based on YOLO11n and includes a backbone feature extraction network, a space-frequency collaborative multi-scale feature aggregation network SFCA-Neck, and a positioning quality dynamic calibration detection head DLQC-Head.

[0019] Infrared targets are typically small in size, have indistinct textures, and are easily obscured by thermal noise in complex backgrounds. Conventional convolutional operations primarily focus on local regions, making it difficult to simultaneously utilize long-distance background information and target edge details. Therefore, this invention incorporates a Global-Local Contextual Feature Enhancement (GLCFE-Block) module deep within the backbone feature extraction network; for example... Figure 2As shown, the Global-Local Context Enhancement (GLCFE-Block) module adopts a structure combining branch connections, cross-stage partial connections, and residual connections, including a cross-stage partial connection context enhancement unit (CSPCE) and a global-local dual-branch attention unit (GLDBA). CSPCE first adjusts the channels of the input deep feature map and divides it along the channel dimension into side-path features and context enhancement features. The side-path features are directly retained without participating in subsequent attention operations, while the context enhancement features are sequentially input into at least one GLDBA for global-local context modeling. Then, the GLDBA output and the side-path features are concatenated along the channel dimension, followed by convolution for feature fusion and channel compression to obtain the context-enhanced deep feature map. In the global branch, GLDBA uses window attention to extract low-frequency background context, while in the local branch, it uses depthwise separable convolution and nonlinear gating to enhance the high-frequency edge response of small targets. The fused features are output via residual connections. Through this structure, the detection network ISTCE-Net can suppress large-area background interference while preserving the local edge information of small infrared targets.

[0020] In target detection tasks, multi-scale feature fusion is particularly important for infrared weak target detection. Shallow feature maps have high spatial resolution and can preserve target edge and positional details; deep feature maps contain richer semantic information and can reflect the contextual relationship between the target and the background. However, traditional feature pyramids typically employ interpolation upsampling, strided convolution downsampling, and simple concatenation operations, which can easily lead to the loss of key pixels of weak targets and cause background thermal noise to propagate between different layers.

[0021] To address the aforementioned issues, this invention proposes the Space-Frequency Cooperative Multi-Scale Feature Aggregation Network (SFCA-Neck), comprising a space-frequency three-branch dynamic calibration module (SFDC-Block) and a channel-shift hybrid feature reconstruction upsampling module (SCMU-Block). Furthermore, when transferring shallow features to deeper layers, a spatial information rearrangement-type downsampling operator is employed to reduce the loss of key pixels of weak targets during the downsampling process. Figure 3 As shown, the space-frequency three-branch dynamic calibration module SFDC-Block adopts a feature partitioning and fusion structure with residual connections. Figure 4 As shown, the Channel Shift Hybrid Feature Reconstruction Upsampling Module (SCMU-Block) is used to restore the spatial resolution of deep feature maps.

[0022] In traditional target detection networks, classification and location regression branches are typically used for prediction. The classification branch determines the target category, while the location regression branch determines the bounding box position. However, the edges of small infrared targets are easily blurred due to thermal diffusion, leading to high uncertainty in bounding box position prediction. This can result in detection results with high classification confidence but inaccurate localization. To address these issues, this invention proposes the DLQC-Head, a dynamic calibration detection head for localization quality. Figure 5 As shown, it includes a classification branch, a location regression branch, and a regression distribution statistical quality assessment unit (RDQA).

[0023] Step 3: Train the detection network ISTCE-Net using the training set data from Step 1; Step 3.1: Input the infrared image to be detected into the backbone feature extraction network, and extract feature maps of different resolutions layer by layer. The output of the backbone feature extraction network includes at least shallow feature maps. Mid-layer feature map and deep feature map ,in Spatial resolution lower than and It also contains richer target semantics and background context information. Step 3.2: The deep feature map output in Step 3.1 As input, global-local context enhancement is performed through the Global-Local Context Feature Enhancement (GLCFE-Block) module. Step 3.1 Output of deep feature map This is used as the input feature map of GLCFE-Block. ,in , , and These represent the height, width, and number of channels of the feature map, respectively. Within CSPCE, first... Divided into bypass features along the channel dimension and context-enhanced features ; Reserved as a cross-stage bypass. Input GLDBA and further divide it into global features along the channel dimension. and local features The data are fed into the global branch and the local branch for parallel processing, respectively.

[0024] Global branch receives global features Attention computation is constructed using local windows as units. Specifically, query features are first obtained through linear mapping. : in This represents the linear mapping matrix of the query features; Again Perform average pooling and obtain key features via linear mapping. Sum value characteristics : in This indicates the average pooling operation performed within the attention window. and These represent linear mapping matrices for key features and value features, respectively; Finally, according to The relevance is used to calculate attention weights, and the relevance is used to calculate attention weights. We perform weighted summation to obtain the global branch output features. : in The channel dimension or scaling factor representing the key feature; Average pooling reduces the computational cost of attention while suppressing high-frequency noise, thus... Low-frequency contextual relationships that characterize a large range of background information.

[0025] Local branches are used to enhance the edge and shape information of small targets. After linear mapping, the local features are input into two parallel depthwise convolutional branches to obtain local response features. and Subsequently, key high-frequency features were screened through element-wise multiplication, nonlinear activation, and gating operations.

[0026] in, This indicates element-wise multiplication. Represents a non-linear activation function. Represents the hyperbolic tangent function. Represents a scaled-down view of local features. This indicates the output characteristics of a local branch.

[0027] Finally, output the global branch. With local branch output The output features are obtained by concatenating along the channel dimension and then using linear mapping and residual connections. In CSPCE, the aforementioned As a context-enhanced branch output, it is compared with bypass features. After being stitched along the channel dimension, Convolution performs channel compression and feature fusion to obtain the CSPCE output features. Then It is fed into subsequent layers of GLCFE-Block to achieve synergistic enhancement of global background information and local edge detail information.

[0028] Step 3.3: The space-frequency collaborative multi-scale feature aggregation network SFCA-Neck receives the shallow feature map output in step 3.1. Mid-layer feature map and the deep feature map enhanced by the GLCFE-Block output in step 3.2. The system performs bidirectional cross-layer feature fusion: In the top-down path, high-level features are upsampled to the same spatial resolution as the next layer's features using SCMU-Block, then concatenated with the current layer's features along the channel dimension and calibrated using SFDC-Block; In the bottom-up path, the fused low-level features are downsampled to the same spatial resolution as the next layer's features using spatial information rearrangement convolution, then concatenated with the current layer's features and calibrated using SFDC-Block, thus outputting a multi-scale feature map that combines high-resolution spatial details and high-level semantic information. In the SFDC-Block three-branch spatial frequency dynamic calibration module, the input features are first processed by convolution and then divided into residual features and features to be calibrated along the channel dimension. The features to be calibrated are input into the global frequency branch, the large kernel spatial branch, and the local detail branch, respectively. The global frequency branch extracts the global energy distribution of the image through fast Fourier transform and inverse fast Fourier transform; the large kernel spatial branch uses 31×1, 1×31, and 31×31 depth convolutions to extract long-distance spatial dependencies in different directions; and the local detail branch uses 1×1 depth convolution to preserve pixel-level local information.

[0029] in, This represents the output characteristics of the global frequency branch. This indicates the output characteristics of the large core spatial branch. This represents the output features of local detail branches. Representing residual characteristics, This represents the calibrated feature map of the SFDC-Block output.

[0030] In the Channel Shift Hybrid Feature Reconstruction Upsampling Module (SCMU-Block), the input features are first upsampled and depthwise separable convolutions, then divided into multiple feature groups along the channel dimension. After spatial shifting and channel blending operations, these feature groups are mapped to the output space using pointwise convolutions. This module promotes information exchange between adjacent regions, reduces feature aliasing caused by conventional upsampling, and enhances the continuity of edges on weak targets.

[0031] In the bottom-up feature transfer process, SFCA-Neck can use a spatial information rearrangement downsampling operator to transfer spatial information in shallow feature maps to the channel dimension, so as to reduce the loss of key pixels of weak targets during the downsampling process.

[0032] In one implementation, the spatial information rearrangement downsampling operator is SPD-Conv; in other implementations, a downsampling structure with the same spatial rearrangement function may also be used.

[0033] Step 3.4: Input the fused multi-scale feature map into the DLQC-Head, a dynamic calibration head for localization quality. The DLQC-Head generates a localization quality calibration value based on the statistical characteristics of the bounding box regression probability distribution, and uses this calibration value to adjust the classification confidence. It then outputs the target category, bounding box position, and corresponding confidence. The classification branch uses two concatenated depthwise separable convolutions and a nonlinear activation function to extract classification features. Let the first... The input feature map for each detection scale is The output of two depthwise separable convolutions is ,in This represents depthwise separable convolution. Represents a non-linear activation function; then through Convolution yields the initial classification score (logit). , its first The value of the channel corresponds to the _th On the first detection scale The initial classification score for the target class. The location regression branch is used to predict the location of the target bounding box. This branch does not directly output a single coordinate value, but instead outputs the offset of each position of the bounding box in a discrete interval. The probability distribution within. Let the output feature of the location regression branch be... Then, the regression probability distribution is obtained through normalization: in, This represents the discrete regression feature of the position regression branch output. Indicates to conduct The regression probability distribution obtained after normalization, where reg_max represents the maximum number of the discrete regression interval.

[0034] When the regression probability distribution is relatively concentrated, it indicates that the location regression branch predicts the bounding box location with relatively certainty and the positioning quality is relatively high; when the regression probability distribution is relatively flat, it indicates that there is a high degree of uncertainty in the bounding box location.

[0035] To reduce additional computational load, the Regression Distribution Statistical Quality Assessment (RDQA) unit does not perform complex entropy calculations, but instead extracts the top k maximum values ​​and their mean from the regression probability distribution: Subsequently, the statistical feature vector Input to the multilayer sensor to generate positioning quality calibration values: Positioning quality calibration value Compared with the initial classification score logit Add them together to obtain the calibrated classification logit: in, This represents the maximum number of responses selected from the regression probability distribution, which can be set according to the requirements of detection accuracy and computational load. express Center front One maximum probability value; express The mean; Indicates by and The resulting location quality statistical feature vector is obtained by splicing. This represents the positioning quality calibration value generated by the MLP; This represents the classification logit after positioning quality calibration. implement The function yields the final classification confidence score used for output.

[0036] Through the above processing, the Regression Distribution Statistical Quality Assessment (RDQA) unit uses the probability distribution statistics from the location regression branch to dynamically calibrate the classification score. When the uncertainty of the bounding box localization result is high, the model can reduce the classification confidence of the corresponding detection result; when the bounding box localization result is relatively reliable, the model can retain the corresponding detection result. Therefore, the Dynamic Quality Calibration Detection Head (DLQC-Head) enables the final classification confidence to simultaneously reflect the target category probability and the bounding box localization quality, improving the reliability of detection result screening and ranking.

[0037] Step 4: Evaluate the performance of the detection network ISTCE-Net using the validation set and optimize the parameters. Step 5: Input the image to be detected into the trained detection network ISTCE-Net to obtain the infrared weak target detection results. Experimental section: To verify the technical effectiveness of this invention, the VEDAI infrared aerial image dataset was used for testing under a set of non-limiting experimental conditions. The VEDAI dataset contains 1246 infrared images, each 1024×1024 pixels, and includes nine categories: cars, trucks, pickups, tractors, RVs, ships, vans, aircraft, and other targets. The dataset is divided into a training set, a test set, and a validation set in a 7:2:1 ratio, containing 873, 249, and 124 images respectively.

[0038] The training parameters were set as follows: input image size 640×640, training epochs 100, batch size 16. The model was trained using random initialization without loading pre-trained weights. The optimizer used was a stochastic gradient descent optimizer with a momentum of 0.937 and a weight decay coefficient of 0.0005. The above-mentioned dataset partitioning, input image size, training epochs, batch size, optimizer type, and parameters are only for illustrating the experimental verification process and do not constitute a necessary limitation for implementing this invention. Table 1 shows the comparison results of different detection methods under the above experimental conditions, illustrating the performance improvement of this invention in a specific experimental scenario.

[0039] Table 1 Comparison Experiments on the VEDAI Dataset As shown in Table 1, the present invention achieves good detection results in all average accuracy indicators. Specifically, the mAP50 reaches 42.55%, which is 4.88 percentage points higher than the basic network YOLO11n; the recall rate reaches 52.09%, which is 6.25 percentage points higher than the basic network; and the mAP50:95 reaches 26.42%.

[0040] Figure 6This diagram illustrates the detection results of the ISTCE-Net detection network on VEDAI infrared aerial imagery test images, including several typical scenes with complex backgrounds such as roads, farmland, and buildings. The bounding boxes in each sub-image represent the output target bounding boxes, and the labels next to the bounding boxes indicate the predicted target category and the classification confidence score after localization quality calibration. Figure 6 As can be seen, this invention can locate and classify small targets such as cars, trucks, and ships in complex backgrounds. Figure 6 This is only used to illustrate the detection output format of the present invention and does not constitute a limitation on the application scenario, target category or detection quantity.

Claims

1. A method for cooperative enhancement detection of weak infrared targets based on improved YOLO11n, characterized in that, Includes the following steps: Step 1: Construct an infrared weak target dataset and divide it into training, testing, and validation sets; Step 2: Construct the infrared weak target collaborative enhancement detection network ISTCE-Net; The detection network ISTCE-Net is based on YOLO11n and includes a backbone feature extraction network, a space-frequency collaborative multi-scale feature aggregation network SFCA-Neck, and a positioning quality dynamic calibration detection head DLQC-Head. The backbone feature extraction network is configured with a deep global-local context feature enhancement module (GLCFE-Block). The GLCFE-Block adopts a structure that combines branch connections, cross-stage partial connections, and residual connections, including a cross-stage partial connection context enhancement unit (CSPCE) and a global-local dual-branch attention unit (GLDBA). The space-frequency collaborative multi-scale feature aggregation network SFCA-Neck includes a space-frequency three-branch dynamic calibration module SFDC-Block and a channel shift hybrid feature reconstruction upsampling module SCMU-Block. The space-frequency three-branch dynamic calibration module SFDC-Block adopts a feature partitioning and fusion structure with residual connections. The positioning quality dynamic calibration detection head DLQC-Head includes a classification branch, a position regression branch, and a regression distribution statistical quality assessment unit RDQA. Step 3: Train the detection network ISTCE-Net using the training set data from Step 1; Step 3.1: Input the infrared image to be detected into the backbone feature extraction network, and extract feature maps of different resolutions layer by layer. The output of the backbone feature extraction network includes at least shallow feature maps. Mid-layer feature map and deep feature map ,in Spatial resolution lower than and Furthermore, it contains richer target semantics and background context information; Step 3.2: The deep feature map output in Step 3.1 As input, global-local context enhancement is performed through the Global-Local Context Feature Enhancement (GLCFE-Block) module; Step 3.3: The space-frequency collaborative multi-scale feature aggregation network SFCA-Neck receives the shallow feature map output in step 3.

1. Mid-layer feature map and the deep feature map enhanced by the GLCFE-Block output in step 3.

2. And perform bidirectional cross-layer feature fusion: In the top-down path, the high-level features are upsampled to the same spatial resolution as the next layer features by SCMU-Block, and then spliced ​​with the features of that layer along the channel dimension and calibrated by SFDC-Block. In the bottom-up path, the fused low-level features are downsampled to the same spatial resolution as the next layer features through spatial information rearrangement convolution, and then concatenated with the features of that layer and calibrated by SFDC-Block to output a multi-scale feature map that has both high-resolution spatial details and high-level semantic information. Step 3.4: Input the fused multi-scale feature map into the localization quality dynamic calibration detection head DLQC-Head. The localization quality dynamic calibration detection head DLQC-Head generates a localization quality calibration value based on the statistical characteristics of the bounding box regression probability distribution, and uses the calibration value to adjust the classification confidence, outputting the target category, bounding box position and corresponding confidence. Step 4: Use the validation set to evaluate the performance of the detection network ISTCE-Net and optimize the parameters; Step 5: Input the image to be detected into the trained detection network ISTCE-Net to obtain the infrared weak target detection results.

2. The infrared weak target cooperative enhancement detection method based on improved YOLO11n according to claim 1, characterized in that, The process in step 3.2 includes: The deep feature map output in step 3.1 This is used as the input feature map of GLCFE-Block. ,in , , and These represent the height, width, and number of channels of the feature map, respectively; within CSPCE, first... Divided into bypass features along the channel dimension and context-enhanced features ; Reserved as a cross-stage bypass. Input GLDBA and further divide it into global features along the channel dimension. and local features The data are fed into the global branch and the local branch for parallel processing, respectively. Global branch receives global features Attention computation is constructed using local windows as units; specifically, query features are first obtained through linear mapping. : in This represents the linear mapping matrix of the query features; Again Perform average pooling and obtain key features via linear mapping. Sum value characteristics : in This indicates the average pooling operation performed within the attention window. and These represent linear mapping matrices for key features and value features, respectively; Finally, according to The relevance is used to calculate attention weights, and the relevance is used to calculate attention weights. We perform weighted summation to obtain the global branch output features. : in The channel dimension or scaling factor representing the key feature; After linear mapping, the local features are input into two parallel depthwise convolutional branches to obtain local response features. and Subsequently, key high-frequency features were screened through element-wise multiplication, nonlinear activation, and gating operations. in, This indicates element-wise multiplication. Represents a non-linear activation function. Represents the hyperbolic tangent function. Represents a scaled-down view of local features. Indicates the output characteristics of local branches; Finally, output the global branch. With local branch output The output features are obtained by concatenating along the channel dimension and then using linear mapping and residual connections. In CSPCE, the aforementioned As a context-enhanced branch output, it is compared with bypass features. After being stitched along the channel dimension, Convolution performs channel compression and feature fusion to obtain the CSPCE output features. Then It is fed into subsequent layers of GLCFE-Block to achieve synergistic enhancement of global background information and local edge detail information.

3. The infrared weak target cooperative enhancement detection method based on improved YOLO11n according to claim 2, characterized in that, The process in step 3.3 includes: In the SFDC-Block spatial frequency three-branch dynamic calibration module, the input features are first processed by convolution and then divided into residual features and features to be calibrated along the channel dimension. The features to be calibrated are input into the global frequency branch, the large kernel spatial branch, and the local detail branch, respectively. The global frequency branch extracts the global energy distribution of the image through fast Fourier transform and inverse fast Fourier transform. The large kernel spatial branch uses 31×1, 1×31, and 31×31 depth convolutions to extract long-distance spatial dependencies in different directions. The local detail branch uses 1×1 depth convolutions to preserve pixel-level local information. in, This represents the output characteristics of the global frequency branch. This indicates the output characteristics of the large core spatial branch. This represents the output features of local detail branches. Representing residual characteristics, This represents the calibrated feature map output by SFDC-Block; In the channel shift hybrid feature reconstruction upsampling module SCMU-Block, the input features are first upsampled and depth-separable convolutional processed, and then divided into multiple feature groups along the channel dimension. After spatial displacement and channel blending operations, different feature groups are mapped to the output space using pointwise convolution. During the bottom-up feature transfer process, the SFCA-Neck employs a spatial information rearrangement downsampling operator to transfer spatial information from shallow feature maps to the channel dimension, thereby reducing the loss of key pixels of weak targets during the downsampling process.

4. The infrared weak target cooperative enhancement detection method based on improved YOLO11n according to claim 3, characterized in that, The spatial information rearrangement downsampling operator is SPD-Conv.

5. The infrared weak target cooperative enhancement detection method based on improved YOLO11n according to claim 4, characterized in that, The process in step 3.4 includes: The classification branch uses two concatenated depthwise separable convolutions and a nonlinear activation function to extract classification features; let the first... The input feature map for each detection scale is The output of two depthwise separable convolutions is ,in This represents depthwise separable convolution. Represents a non-linear activation function; then through Convolution yields the initial classification score. , its first The value of the channel corresponds to the _th On the first detection scale Initial classification score for the target class; The location regression branch does not directly output a single coordinate value, but instead outputs the offset of each position of the bounding box in a discrete interval. The probability distribution within; Let the output feature of the location regression branch be Then, the regression probability distribution is obtained through normalization: in, This represents the discrete regression feature of the position regression branch output. Indicates to conduct The regression probability distribution obtained after normalization, where reg_max represents the maximum index of the discrete regression interval; The regression distribution statistical quality assessment unit (RDQA) does not perform complex entropy calculations, but instead extracts the top k maximum values ​​and their mean from the regression probability distribution: Subsequently, the statistical feature vector Input to the multilayer sensor to generate positioning quality calibration values: Positioning quality calibration value Compared with the initial classification score logit Add them together to obtain the calibrated classification logit: in, This represents the maximum number of responses selected from the regression probability distribution, which can be set according to the requirements of detection accuracy and computational load. express Center front One maximum probability value; express The mean; Indicates by and The resulting location quality statistical feature vector is obtained by splicing. This represents the positioning quality calibration value generated by the MLP; This represents the classification logit after positioning quality calibration. implement The function yields the final classification confidence score used for output.

Images