Self-adaptive full-scale infrared target detection network based on YOLO

Abstract

The invention relates to the technical field of infrared target detection, and discloses a YOL0-based adaptive full-scale infrared target detection network, which comprises a trunk feature extraction network, a neck feature fusion network, a detection head network and a training optimization module, and is characterized in that all the modules are sequentially connected in series to form a complete detection link; the infrared image multi-dimensional feature extraction system is used for infrared image multi-dimensional feature extraction and comprises a convolution layer, an SPPF module and a C2MFE module which are connected in sequence, and the C2MFE module replaces a standard convolution layer in a traditional C2f module through multi-kernel feature extraction convolution (MFEEConv) to achieve multi-direction and full-scale feature capture; and the neck feature fusion network is connected with the output end of the trunk feature extraction network, comprises a multi-scale feature fusion module (MFFM) and a feature pyramid structure, and is used for enhancing the feature correlation of different levels. The adaptive full-scale infrared target detection network based on YOL0 can efficiently adapt to complex scenes such as low illumination and severe weather, and realizes cross-scene full-scale infrared target accurate detection.

Description

Technical Field

[0001] This invention relates to the field of infrared target detection technology, specifically to an adaptive full-scale infrared target detection network based on YOLO. Background Technology

[0002] Infrared target detection plays an irreplaceable role in many key fields due to its unique perception capabilities in complex environments such as low light and harsh weather. However, infrared images have inherent drawbacks such as low contrast, blurred details, significant noise interference, and large differences in target scale, which pose numerous challenges to traditional detection algorithms. Existing algorithms are mostly optimized for single-scale targets. While high-sensitivity feature extraction is needed to capture subtle features for small targets at long distances, stronger global semantic understanding is required for large targets at close range. Traditional algorithms struggle to balance these two aspects, leading to issues such as missed detections of small targets and local overfitting of large targets. Furthermore, traditional convolutional modules often use single-directional kernels, resulting in limited receptive fields and difficulty in capturing the multi-directional distribution features of infrared targets. Additionally, the low signal-to-noise ratio of infrared images makes feature extraction susceptible to background noise, diluting key features. Traditional feature fusion methods often involve simple concatenation or element-wise addition, failing to fully exploit the correlations between features at different levels. This results in a mismatch between shallow spatial details and deep semantic information, impacting detection accuracy. Existing algorithms lack generalization ability when facing complex scenarios, leading to a significant drop in detection accuracy. Most high-precision algorithms rely on complex network structures with large parameter and computational requirements, making them difficult to deploy in resource-constrained scenarios such as drones and edge devices. Lightweight algorithms, on the other hand, often sacrifice detection accuracy, failing to meet practical application needs. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an adaptive full-scale infrared target detection network based on YOLO. This network has the advantages of being able to efficiently adapt to complex scenarios such as low light and severe weather, and achieving accurate detection of infrared targets across all scenarios and scales. It solves the problems of insufficient generalization ability and significant drop in detection accuracy when facing different complex scenarios. Most high-precision algorithms rely on complex network structures, with a large number of parameters and computational load, making them difficult to deploy in resource-constrained scenarios such as drones and edge devices. On the other hand, lightweight algorithms often sacrifice detection accuracy and cannot meet the needs of practical applications.

[0004] (II) Technical Solution To achieve the above objectives, this invention provides the following technical solution: an adaptive full-scale infrared target detection network based on YOLO, comprising a backbone feature extraction network, a neck feature fusion network, a detection head network, and a training and optimization module, with each module connected in series to form a complete detection link: The backbone feature extraction network is used for multi-dimensional feature extraction of infrared images. It includes a convolutional layer, an SPPF module and a C2MFE module connected in sequence. The C2MFE module replaces the standard convolutional layer in the traditional C2f module with a multi-kernel feature extraction convolution (MFEConv) to achieve multi-directional, full-scale feature capture. The neck feature fusion network, connected to the output of the backbone feature extraction network, includes a multi-scale feature fusion module (MFFM) and a feature pyramid structure, which are used to enhance the correlation of features at different levels and achieve efficient fusion of shallow spatial details and deep semantic information. The detection head network, connected to the output of the neck feature fusion network, includes a cross-attention fusion module (CAFM) and a classification and regression branch. The cross-attention fusion module (CAFM) is deployed before the output of the feature pyramid layers P3, P4, and P5, and is used for deep and shallow feature cross-interaction and noise suppression. The training optimization module is used to optimize parameters during the model training process. It includes an adaptive data augmentation unit and a hybrid loss function unit to improve the model's generalization ability and detection accuracy.

[0005] Preferably, the multi-kernel feature extraction convolution (MFEConv) includes a channel shift preprocessing unit, a heterogeneous convolutional group, a joint attention mechanism module (FAM), and a residual connection unit: The channel shifting preprocessing unit shifts and superimposes the input feature map in multiple directions to enhance feature diversity. The heterogeneous convolutional group divides the preprocessed features into 6 groups according to channels, and performs parallel convolutions using 3×3 convolutional kernels in 6 directions: horizontal left and right, vertical up and down, and left and right diagonal, to increase the receptive field. The Joint Attention Mechanism (FAM) module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule focuses the main channel through thermal weighted average pooling and inverse weighted max pooling, while the spatial attention submodule enhances the details of the infrared target region. The residual connection unit achieves residual fusion between the convolution output and the input features through learnable scaling parameters, thereby improving the stability of feature propagation.

[0006] Preferably, the multi-scale feature fusion module (MFFM) includes a feature preprocessing unit, a three-branch feature processing unit, and a residual fusion unit: The feature preprocessing unit uses a 1×1 convolutional layer to uniformly input the number of feature channels and generates global features through element-wise summation; The three-branch feature processing unit: the first branch introduces a channel attention mechanism (CBAM) to focus on key channel features; the second branch generates filtering coefficients through sigmoid activation, and then concatenates and fuses them after grouping convolution to capture channel correlation; the third branch separates strong / weak feature sets through spatial information weights, and then aggregates them after processing by depthwise separable convolution and 1×1 convolution respectively. The residual fusion unit performs residual connection on the three-branch output features and introduces efficient local attention (ELA) to suppress noise, outputting fused features.

[0007] Preferably, the cross-attention fusion module (CAFM) includes a cross-interactive attention unit (CA) and a dynamic gating unit (DG): The cross-interactive attention unit (CA) generates a query matrix of deep and shallow features through the projection layer, and generates cross-branch attention weights through matrix multiplication and the Softmax function to achieve complementary enhancement of deep and shallow features; The Dynamic Gating Unit (DG) concatenates the original features with cross-attention enhanced features, captures the local context through 5×5 convolution, generates an adaptive weight matrix, dynamically adjusts the contribution ratio of deep and shallow features, and suppresses background noise.

[0008] Preferably, the adaptive data augmentation unit of the training optimization module includes infrared image-specific augmentation strategies: adaptive histogram equalization based on the low contrast characteristics of infrared images, random scaling and cropping for target scale differences, noise addition and fog effect simulation to simulate severe weather; the hybrid loss function unit adopts a weighted combination of classification loss (Focal Loss) and regression loss (CIoU Loss), and the weight coefficients are dynamically adjusted according to the training iteration process.

[0009] Preferably, the six convolutional kernels of the heterogeneous convolutional group correspond to the horizontal left direction, horizontal right direction, vertical up direction, vertical down direction, left-up-right-down diagonal direction, and right-up-left-down diagonal direction, respectively. After each convolution, it is processed by batch normalization (BN) and SiLU activation function in sequence, and the parallel output features are input into the joint attention mechanism module after channel concatenation.

[0010] Preferably, the strong / weak feature set separation process of the third branch includes: generating spatial information weights through 1×1 convolution and BN-Sigmoid activation, obtaining feature weight thresholds through adaptive average pooling, comparing the spatial information weights with the threshold using a threshold function to generate an attention map, and separating the strong feature set and the weak feature set through element-wise multiplication.

[0011] Preferably, the projection layer of the cross-interactive attention unit (CA) is transformed into a lightweight structure using depthwise separable convolution, and a temperature scaling factor is introduced during matrix multiplication to optimize the weight distribution and avoid gradient instability.

[0012] Preferably, the input image size of the backbone feature extraction network is 640×640. The number of image channels is converted from 1 channel (infrared single channel) to 32 channels through a convolutional layer. After the feature dimension is compressed by the SPPF module, it is input into the C2MFE module. The number of C2MFE modules is set to 3 to 5 according to the network depth requirements.

[0013] Preferably, the classification and regression branch of the detection head network adopts the detection head structure of YOLOv10, outputs the target category, confidence score and bounding box coordinates, filters redundant detection boxes through non-maximum suppression (NMS), and outputs the final detection result.

[0014] (III) Beneficial Effects Compared with existing technologies, this invention provides an adaptive full-scale infrared target detection network based on YOLO, which has the following advantages: 1. This YOLO-based adaptive full-scale infrared target detection network increases the receptive field through heterogeneous convolutional groups of MFEConv, hierarchical feature processing of MFFM, and deep and shallow feature fusion of CAFM, achieving full-scale coverage detection from small targets at long distances to large targets at close distances.

[0015] 2. This YOLO-based adaptive full-scale infrared target detection network captures multi-directional features through MFEConv, enhances feature correlation through MFFM, achieves feature complementarity and noise suppression through CAFM, and significantly improves feature expression capability through a three-level feature optimization mechanism, resulting in improved detection accuracy compared to YOLOv10n.

[0016] 3. This YOLO-based adaptive full-scale infrared target detection network improves scene adaptability through ELA attention, dynamic gating to suppress infrared noise, and adaptive data augmentation. Its cross-scene deployment capability has been verified on the HIT-UAV and IRSTD-1k cross-scene datasets.

[0017] 4. This YOLO-based adaptive full-scale infrared target detection network has only 2.56M network parameters and 8.2G FLOPs, which is far lower than mainstream detection models. It can be efficiently deployed in resource-constrained scenarios such as drones and edge devices. Attached Figure Description

[0018] Figure 1 This is a flowchart of the entire detection process during the inference stage of the YOLO-based adaptive full-scale infrared target detection network proposed in this invention. Figure 2 This is a flowchart of the training, optimization, and iterative process for the YOLO-based adaptive full-scale infrared target detection network proposed in this invention. Figure 3The flowchart of multi-kernel feature extraction of the adaptive full-scale infrared target detection network based on YOLO proposed in this invention; Figure 4 The present invention presents a flowchart of feature fusion and accurate detection of an adaptive full-scale infrared target detection network based on YOLO. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1-4 The adaptive full-scale infrared target detection network based on YOLO includes a backbone feature extraction network for multi-dimensional feature extraction of infrared images. It includes a convolutional layer, an SPPF module, and a C2MFE module connected in sequence. The C2MFE module replaces the standard convolutional layer in the traditional C2f module with a multi-kernel feature extraction convolution (MFEConv) to achieve multi-directional, full-scale feature capture. The neck feature fusion network, connected to the output of the backbone feature extraction network, includes a multi-scale feature fusion module (MFFM) and a feature pyramid structure, which are used to enhance the correlation of features at different levels and achieve efficient fusion of shallow spatial details and deep semantic information. The detection head network, connected to the output of the neck feature fusion network, includes a cross-attention fusion module (CAFM) and a classification and regression branch. The cross-attention fusion module (CAFM) is deployed before the output of the feature pyramid layers P3, P4, and P5, and is used for deep and shallow feature cross-interaction and noise suppression. The training optimization module is used to optimize parameters during the model training process. It includes an adaptive data augmentation unit and a hybrid loss function unit to improve the model's generalization ability and detection accuracy.

[0021] Preferably, the multi-kernel feature extraction convolution (MFEConv) includes a channel shift preprocessing unit, a heterogeneous convolutional group, a joint attention mechanism module (FAM), and a residual connection unit: The channel shifting preprocessing unit shifts and superimposes the input feature map in multiple directions to enhance feature diversity. The heterogeneous convolutional group divides the preprocessed features into 6 groups according to channels, and performs parallel convolutions using 3×3 convolutional kernels in 6 directions: horizontal left and right, vertical up and down, and left and right diagonal, to increase the receptive field. The Joint Attention Mechanism (FAM) module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule focuses the main channel through thermal weighted average pooling and inverse weighted max pooling, while the spatial attention submodule enhances the details of the infrared target region. The residual connection unit achieves residual fusion between the convolution output and the input features through learnable scaling parameters, thereby improving the stability of feature propagation.

[0022] Specifically, the principle of multi-directional feature enhancement and residual propagation is employed. Channel shifting enriches feature diversity, heterogeneous convolutional groups expand the receptive field to capture multi-directional features, the FAM attention mechanism double-screens key features, and residual connections ensure the stability of feature propagation. The core solution addresses the problems of single feature extraction and vanishing gradients in traditional convolutional methods.

[0023] Preferably, the multi-scale feature fusion module (MFFM) includes a feature preprocessing unit, a three-branch feature processing unit, and a residual fusion unit: The feature preprocessing unit uses a 1×1 convolutional layer to uniformly input the number of feature channels and generates global features through element-wise summation; The three-branch feature processing unit: the first branch introduces a channel attention mechanism (CBAM) to focus on key channel features; the second branch generates filtering coefficients through sigmoid activation, and then concatenates and fuses them after grouping convolution to capture channel correlation; the third branch separates strong / weak feature sets through spatial information weights, and then aggregates them after processing by depthwise separable convolution and 1×1 convolution respectively. The residual fusion unit performs residual connection on the three-branch output features and introduces efficient local attention (ELA) to suppress noise, outputting fused features.

[0024] Specifically, the principle of hierarchical feature fusion and noise suppression is as follows: multi-dimensional feature optimization is achieved through three-branch parallel processing. CBAM focuses on key channels, grouped convolution enhances channel correlation, and strong / weak feature separation and aggregation improves feature quality. Then, residual fusion and ELA attention suppress noise. The core is to solve the problems of poor performance and severe noise interference in traditional feature fusion.

[0025] Preferably, the cross-attention fusion module (CAFM) includes a cross-interactive attention unit (CA) and a dynamic gating unit (DG): The cross-interactive attention unit (CA) generates a query matrix of deep and shallow features through the projection layer, and generates cross-branch attention weights through matrix multiplication and the Softmax function to achieve complementary enhancement of deep and shallow features; The Dynamic Gating Unit (DG) concatenates the original features with cross-attention enhanced features, captures the local context through 5×5 convolution, generates an adaptive weight matrix, dynamically adjusts the contribution ratio of deep and shallow features, and suppresses background noise.

[0026] Specifically, the principle of deep and shallow feature complementarity and dynamic adaptation is employed. By establishing the association between deep and shallow features through cross-interactive attention, and by dynamically adjusting the feature contribution ratio using a gating unit, the system achieves both feature complementarity enhancement and adaptation to the needs of target detection at different scales. The core objective is to address the problems of feature fragmentation between deep and shallow layers and insufficient noise suppression.

[0027] Preferably, the adaptive data augmentation unit of the training optimization module includes infrared image-specific augmentation strategies: adaptive histogram equalization based on the low contrast characteristics of infrared images, random scaling and cropping for target scale differences, noise addition and fog effect simulation to simulate severe weather; the hybrid loss function unit adopts a weighted combination of classification loss (Focal Loss) and regression loss (CIoU Loss), and the weight coefficients are dynamically adjusted according to the training iteration process.

[0028] Specifically, the principle of scene adaptation and loss balance optimization is as follows: Complex scenes are simulated using infrared-specific data enhancement to expand sample diversity; a hybrid loss function dynamically balances the priority of classification and regression optimization, the core of which is to solve the problems of weak model generalization ability, class imbalance, and low bounding box regression accuracy.

[0029] Preferably, the six convolutional kernels of the heterogeneous convolutional group correspond to the horizontal left direction, horizontal right direction, vertical up direction, vertical down direction, left-up-right-down diagonal direction, and right-up-left-down diagonal direction, respectively. After each convolution, it is processed by batch normalization (BN) and SiLU activation function in sequence, and the parallel output features are input into the joint attention mechanism module after channel concatenation.

[0030] Specifically, the principle of multi-directional convolution parallelism and feature normalization is employed. Six convolutional kernels comprehensively cover the target feature distribution. After convolution, features are normalized through Batch Normalization (BN) and SiLU. The core principle is to ensure the integrity of multi-directional feature extraction and training stability.

[0031] Preferably, the strong / weak feature set separation process of the third branch includes: generating spatial information weights through 1×1 convolution and BN-Sigmoid activation, obtaining feature weight thresholds through adaptive average pooling, comparing the spatial information weights with the threshold using a threshold function to generate an attention map, and separating the strong feature set and the weak feature set through element-wise multiplication.

[0032] Specifically, the principle of feature weight selection and hierarchical optimization involves generating spatial information weights and comparing them with thresholds to separate strong and weak feature sets and optimize them accordingly. The core is to achieve accurate feature selection and hierarchical enhancement, thereby increasing the proportion of effective features.

[0033] Preferably, the projection layer of the cross-interactive attention unit (CA) is transformed into a lightweight structure using depthwise separable convolution, and a temperature scaling factor is introduced during matrix multiplication to optimize the weight distribution and avoid gradient instability.

[0034] Specifically, the principles of lightweight projection and gradient stabilization optimization are employed. Depthwise separable convolution reduces the computational cost of the projection layer, and a temperature scaling factor optimizes the weight distribution. The core goal is to achieve lightweight design and improved training stability while ensuring effective feature transformation.

[0035] Preferably, the input image size of the backbone feature extraction network is 640×640. The number of image channels is converted from 1 channel (infrared single channel) to 32 channels through a convolutional layer. After the feature dimension is compressed by the SPPF module, it is input into the C2MFE module. The number of C2MFE modules is set to 3 to 5 according to the network depth requirements.

[0036] Specifically, the principles of input standardization and feature extraction adaptation are as follows: The input image size and number of channels are standardized; the SPPF module compresses dimensions while preserving global information; the number of C2MFE modules is configurable. The core principle is to ensure the consistency of input features and the flexible adaptation of feature extraction capabilities.

[0037] Preferably, the classification and regression branch of the detection head network adopts the detection head structure of YOLOv10, outputs the target category, confidence score and bounding box coordinates, filters redundant detection boxes through non-maximum suppression (NMS), and outputs the final detection result.

[0038] Specifically, the principle of accurate output and redundancy filtering is employed. A mature detection head structure ensures output efficiency, and the NMS algorithm filters redundant detection frames. The core objective is to ensure the accuracy and simplicity of the detection results, avoiding invalid output interference.

[0039] Based on the core logic of feature hierarchical optimization and full-link collaborative adaptation, this invention systematically solves the pain points in infrared target detection, such as poor multi-scale adaptation, insufficient feature extraction, large background noise interference, and weak cross-scene generalization ability, through the linkage and cooperation of four major modules, and achieves accurate detection in all scenarios.

[0040] To address the scale difference between infrared targets that are small at long distances and large at close distances, this paper proposes a method that uses heterogeneous convolutional groups to increase the receptive field of features, employs a feature pyramid structure to capture features of small, medium, and large-scale targets, and combines this with the cross-fusion of shallow and deep features from the CAFM module. This allows the network to adapt to the detection needs of targets at different scales, achieving full-scale coverage. A four-level feature optimization chain is constructed, from channel shift preprocessing to multi-directional feature extraction, attention filtering, and residual fusion. Channel shift and heterogeneous convolutional groups enrich feature diversity and expand the receptive field; attention mechanisms such as FAM, CBAM, and ELA filter key features and suppress redundant noise; and residual fusion ensures the stability of feature propagation, ultimately enhancing the effective features of the target and suppressing infrared image noise.

[0041] Adaptive data augmentation simulates different scenarios such as low light and severe weather, expanding the diversity of training samples. A hybrid loss function dynamically balances classification and regression accuracy, allowing the model to fully learn image features from different scenarios during training, improving its adaptability when deployed across scenarios. Lightweight designs such as depthwise separable convolution, channel splitting, and attention filtering are adopted in the core modules (MFEConv, CAFM). While reducing the number of parameters (only 2.56M) and computational cost (8.2G FLOPs), feature hierarchical optimization and multi-module collaboration ensure that detection accuracy is not reduced, achieving a balance between lightweight deployment and high detection accuracy.

[0042] The backbone feature extraction network extracts multi-directional and multi-scale raw features from infrared images; the neck feature fusion network performs deep and shallow fusion and enhancement of the raw features to generate more expressive fused features; the detection head network performs targeted optimization of the fused features (noise suppression, feature complementarity), and then outputs detection results through a classification and regression branch, forming a closed-loop feature flow link of extraction, fusion, optimization, and detection; during the training phase, the adaptive data augmentation unit expands the sample diversity, and the hybrid loss function optimizes the parameters of each module through backpropagation; during the inference phase, the optimized model parameters are reused to quickly execute the preprocessing, feature extraction, feature fusion, detection, and result output process, ensuring that the feature extraction and detection capabilities learned during training are efficiently implemented in practical applications.

[0043] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A YOLO-based adaptive full-scale infrared target detection network, characterized in that, It includes a backbone feature extraction network, a neck feature fusion network, a detection head network, and a training and optimization module. These modules are connected in series to form a complete detection chain. The backbone feature extraction network is used for multi-dimensional feature extraction of infrared images. It includes a convolutional layer, an SPPF module and a C2MFE module connected in sequence. The C2MFE module replaces the standard convolutional layer in the traditional C2f module with a multi-kernel feature extraction convolution (MFEConv) to achieve multi-directional, full-scale feature capture. The neck feature fusion network, connected to the output of the backbone feature extraction network, includes a multi-scale feature fusion module (MFFM) and a feature pyramid structure, which are used to enhance the correlation of features at different levels and achieve efficient fusion of shallow spatial details and deep semantic information. The detection head network, connected to the output of the neck feature fusion network, includes a cross-attention fusion module (CAFM) and a classification and regression branch. The cross-attention fusion module (CAFM) is deployed before the output of the feature pyramid layers P3, P4, and P5, and is used for deep and shallow feature cross-interaction and noise suppression. The training optimization module is used to optimize parameters during the model training process. It includes an adaptive data augmentation unit and a hybrid loss function unit to improve the model's generalization ability and detection accuracy.

2. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The multi-kernel feature extraction convolution (MFEConv) includes a channel shift preprocessing unit, a heterogeneous convolutional group, a joint attention mechanism module (FAM), and a residual connection unit: The channel shifting preprocessing unit shifts and superimposes the input feature map in multiple directions to enhance feature diversity. The heterogeneous convolutional group divides the preprocessed features into 6 groups according to channels, and performs parallel convolutions using 3×3 convolutional kernels in 6 directions: horizontal left and right, vertical up and down, and left and right diagonal, to increase the receptive field. The Joint Attention Mechanism (FAM) module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule focuses the main channel through thermal weighted average pooling and inverse weighted max pooling, while the spatial attention submodule enhances the details of the infrared target region. The residual connection unit achieves residual fusion between the convolution output and the input features through learnable scaling parameters, thereby improving the stability of feature propagation.

3. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The multi-scale feature fusion module (MFFM) includes a feature preprocessing unit, a three-branch feature processing unit, and a residual fusion unit. The feature preprocessing unit uses a 1×1 convolutional layer to uniformly input the number of feature channels and generates global features through element-wise summation; Three-branch feature processing unit: The first branch introduces a channel attention mechanism (CBAM) to focus on key channel features; The second branch generates filtering coefficients through Sigmoid activation, and then concatenates and merges them after grouping convolution to capture channel correlation. The third branch separates strong / weak feature sets using spatial information weights, and then aggregates them after processing them with depthwise separable convolution and 1×1 convolution, respectively. The residual fusion unit performs residual connection on the three-branch output features and introduces efficient local attention (ELA) to suppress noise, outputting fused features.

4. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The Cross-Attention Fusion Module (CAFM) includes a Cross-Interactive Attention Unit (CA) and a Dynamic Gating Unit (DG): The cross-interactive attention unit (CA) generates a query matrix of deep and shallow features through the projection layer, and generates cross-branch attention weights through matrix multiplication and the Softmax function to achieve complementary enhancement of deep and shallow features; The Dynamic Gating Unit (DG) concatenates the original features with cross-attention enhanced features, captures the local context through 5×5 convolution, generates an adaptive weight matrix, dynamically adjusts the contribution ratio of deep and shallow features, and suppresses background noise.

5. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The adaptive data augmentation unit of the training optimization module includes infrared image-specific augmentation strategies: adaptive histogram equalization based on the low contrast characteristics of infrared images, random scaling and cropping for target scale differences, noise addition and fog effect simulation to simulate severe weather; the hybrid loss function unit adopts a weighted combination of classification loss (Focal Loss) and regression loss (CIoU Loss), and the weight coefficients are dynamically adjusted according to the training iteration process.

6. The YOLO-based adaptive full-scale infrared target detection network according to claim 2, characterized in that: The six convolutional kernels of the heterogeneous convolutional group correspond to the horizontal left direction, horizontal right direction, vertical up direction, vertical down direction, left-up-right-down diagonal direction, and right-up-left-down diagonal direction, respectively. After each convolution, it is processed by batch normalization (BN) and SiLU activation function in sequence. The parallel output features are concatenated by channels and then input into the joint attention mechanism module.

7. The YOLO-based adaptive full-scale infrared target detection network according to claim 3, characterized in that: The strong / weak feature set separation process of the third branch includes: generating spatial information weights through 1×1 convolution and BN-Sigmoid activation, obtaining feature weight thresholds through adaptive average pooling, comparing spatial information weights with the threshold using a threshold function to generate an attention map, and separating strong and weak feature sets through element-wise multiplication.

8. The YOLO-based adaptive full-scale infrared target detection network according to claim 4, characterized in that: The projection layer of the cross-interactive attention unit (CA) is made lightweight by using depthwise separable convolution. A temperature scaling factor is introduced during the matrix multiplication calculation to optimize the weight distribution and avoid gradient instability.

9. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The input image size of the backbone feature extraction network is 640×640. The number of image channels is converted from 1 channel (single infrared channel) to 32 channels through convolutional layers. After the feature dimension is compressed by the SPPF module, it is input into the C2MFE module. The number of C2MFE modules is set to 3 to 5 according to the network depth requirements.

10. The YOLO-based adaptive full-scale infrared target detection network according to claim 1, characterized in that: The classification and regression branch of the detection head network adopts the detection head structure of YOLOv10, outputting the target category, confidence score, and bounding box coordinates. Redundant detection boxes are filtered through non-maximum suppression (NMS) to output the final detection result.

Images