Infrared image target detection method and device, computer equipment and storage medium

Abstract

The invention relates to an infrared image target detection method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring a multi-scale feature map through a feature extraction network, and realizing convolution kernel real-time selection and feature extraction adaptive adjustment by fusing an adaptive weight division unit adopted by the multi-scale feature map with a dynamic convolution kernel weighting mechanism. A deep scale feature map in the multi-scale feature map is processed by a polarization dynamic attention fusion unit, and the polarization dynamic attention fusion unit enhances and optimizes deep features by means of multi-scale convolution, an attention mechanism and dynamic feature regulation and control. And inputting the processed deep feature map and other multi-scale feature maps into an adaptive hierarchical feature path aggregation network, carrying out cross-scale fusion from top to bottom and from bottom to top, and carrying out fusion optimization with the original feature maps of all scales through an adaptive hierarchical feature fusion unit to obtain an optimized fusion feature map. And finally, the detection head completes target detection based on the image, so that the efficiency and accuracy of infrared image target detection are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of infrared target detection, in particular to an infrared image target detection method and device, computer equipment and storage medium. BACKGROUND

[0002] In the task of target detection in complex environments, infrared imaging technology shows significant advantages due to its unique thermal radiation perception capability. Unlike visible light imaging, infrared systems generate thermal distribution maps by capturing surface temperature differences, making them indispensable in applications such as night surveillance, military reconnaissance, medical diagnosis, and industrial inspection.

[0003] Infrared target detection faces multiple technical challenges, such as targets often exhibiting low signal-to-noise ratio characteristics, shape and texture information being blurred; complex backgrounds including clouds, sea waves, or urban heat island effects often appear with target thermal radiation aliasing; and noise interference further exacerbates the difficulty of feature extraction. These characteristics make traditional edge detection or texture analysis methods directly unsuitable. Currently, deep learning technology is a key solution to the accuracy and speed challenges in infrared target detection, and models such as Faster R-CNN, YOLO series, and Transformer-based architectures (such as DETR, DINO, RT-DETR, etc.) are widely used in this field, and are constantly being optimized, such as improving small target detection performance, reducing computational cost, and achieving real-time detection.

[0004] However, the characteristics of infrared images have special requirements for algorithms. Existing related algorithms have their own limitations, such as InfraVision-FusionNet having a delay in detecting high-speed targets in foggy conditions, multi-scale complementary feature fusion strategies having poor real-time performance in high dynamic aerial scenes, and some algorithms having deficiencies in computational efficiency, adaptability to complex environments, dense or overlapping target detection, etc. SUMMARY

[0005] Therefore, it is necessary to provide an infrared image target detection method, device, computer equipment and storage medium that can effectively improve target detection efficiency and accuracy in view of the above technical problems.

[0006] An infrared image target detection method, the method comprising: obtaining an infrared image to be subjected to target detection; inputting the infrared image into a feature extraction network to extract features and obtain a multi-scale feature map, wherein the feature extraction network uses an adaptive weight division unit to extract features, and the adaptive weight division unit realizes real-time selection and adaptive adjustment of convolution kernels for feature extraction by fusing a dynamic convolution kernel weighting mechanism. The deep layer scale feature map in the multi-scale feature map is processed by using a polarization dynamic attention fusion unit to obtain a processed deep layer scale feature map. The polarization dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep layer scale feature map through multi-scale convolution, attention mechanism and dynamic feature regulation. The processed deep layer scale feature map and other multi-scale feature maps are input into an adaptive hierarchical feature path aggregation network, cross-scale fusion is performed on each scale feature map of each input from top to bottom and from low to high, and the result of the cross-scale fusion is fused and optimized with each scale feature map of the input by using an adaptive hierarchical feature fusion unit to obtain an optimized fused feature map. A detection head is used to detect a target according to the optimized fused feature map to obtain a detection result.

[0007] In one of the embodiments, in the adaptive weight division unit: The input image is respectively split into two parts and input into two parallel adaptive weight division convolution branches. In the adaptive weight division convolution branch, adaptive average pooling, dynamic kernel weight generation and weighted convolution are used to extract differentiated features from the input data, and then the features are fused by concatenation to enhance feature expression. The data output by the two parallel adaptive weight division convolution branches is concatenated and then passed through a convolution layer to obtain a feature map of a certain scale.

[0008] In one of the embodiments, the feature extraction network includes a convolution layer and a plurality of feature extraction blocks composed of a convolution layer and the adaptive weight division unit. The input image is sequentially input into a plurality of feature extraction blocks after passing through a convolution layer, and different scale feature maps are output by the adaptive weight division units in the feature extraction blocks.

[0009] In one of the embodiments, the polarization dynamic attention fusion unit takes a deep layer scale feature map and corresponding position encoding as dual-channel input, first uses a Polo linear attention mechanism to perform long-range association, uses an ADD block to realize feature accumulation, uses a DyT block to dynamically regulate features, uses a reinforcement channel attention block to focus on key channel features, and finally uses a Mona block for integration.

[0010] In one of the embodiments, in the Mona block: The input data is first normalized and then dimensionally reduced by a Down Projection layer. The dimensionally reduced data is divided into multiple paths, and different depth convolution layers are used to extract corresponding scale features. After the features of multiple scales are fused by an Average layer, residual fusion is performed with the dimensionally reduced data. The data obtained after fusion is further subjected to a convolution layer, a GeLU activation layer, and an Up Projection layer for dimensionality increase, wherein the data after the convolution layer is also subjected to intermediate residual fusion with the data after dimensionality reduction; The data after dimensionality increase is subjected to residual fusion with the input data, completing multi-scale feature depth integration and enhancement.

[0011] In one of the embodiments, the reinforced channel attention block is connected after the Mona block, and in the reinforced channel attention block: The features obtained after the Mona block and the features not subjected to the Mona block are respectively processed in a double-branch processing mode, wherein the features obtained after the Mona block are subjected to regular feature transformation through one branch, and the features not subjected to the Mona block are subjected to multi-scale convolution and up-sampling to mine features through another branch; After the output features of the two branches are spliced and fused, the features are intelligently weighted by using a gating mechanism, realizing fine integration and enhancement of the features.

[0012] In one of the embodiments, in the adaptive hierarchical feature fusion unit: Two input features are respectively subjected to dimensionality reduction to obtain first reduced dimension features and second reduced dimension features; After the first reduced dimension features and the second reduced dimension features are added, a convolution layer is used to realize preliminary fusion to generate baseline features; The first reduced dimension features and the second reduced dimension features are respectively subjected to optimization processing by a local-global attention block to obtain first optimized features and second optimized features, wherein the local-global attention block includes two parallel local and global attention branches; The baseline features, the first optimized features, and the second optimized features are spliced and fused, and then sequentially pass through a composite block composed of a convolution layer, a reparameterization convolution layer, and a convolution layer (Conv1x1, RepConv3x3, Conv1x1) to obtain output features of the adaptive hierarchical feature fusion unit.

[0013] The application also provides an infrared image target detection device, which comprises: An image acquisition module is configured to acquire an infrared image to be subjected to target detection. A multi-scale feature extraction module is configured to input the infrared image into a feature extraction network to extract features and obtain a multi-scale feature map, wherein the feature extraction network adopts an adaptive weight division unit to extract features, and the adaptive weight division unit realizes real-time selection and adaptive adjustment of convolution kernels by using a fusion dynamic convolution kernel weighting mechanism. The deep layer scale feature map optimization module is configured to process a deep layer scale feature map in the multi-scale feature map by using a polarization dynamic attention fusion unit to obtain a processed deep layer scale feature map, wherein the polarization dynamic attention fusion unit is configured to realize feature enhancement and optimization processing of the deep layer scale feature map by using multi-scale convolution, attention mechanism and dynamic feature regulation. The multi-scale feature fusion module is configured to input the processed deep layer scale feature map and other multi-scale feature maps into an adaptive hierarchical feature path aggregation network, perform top-down and low-to-up cross-scale fusion on each scale feature map of each input, and fuse and optimize the cross-scale fusion result and each scale feature map of each input by using an adaptive hierarchical feature fusion unit to obtain an optimized fusion feature map. The target detection module is configured to perform target detection according to the optimized fusion feature map by using a detection head to obtain a detection result.

[0014] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the specific steps in the above infrared image target detection method when executing the computer program.

[0015] A computer readable storage medium stores a computer program, and the computer program implements the specific steps in the above infrared image target detection method when executed by a processor.

[0016] The above infrared image target detection method, device, computer device and storage medium, by inputting an infrared image into a feature extraction network to extract features to obtain a multi-scale feature map, wherein the adaptive weight division unit is used for feature extraction in the feature extraction network, the adaptive weight division unit realizes real-time selection and adaptive adjustment of the convolution kernel by using the dynamic convolution kernel weighting mechanism, and realizes feature extraction, the deep layer scale feature map in the multi-scale feature map is processed by using the polarization dynamic attention fusion unit to obtain the processed deep layer scale feature map, the polarization dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep layer scale feature map by using multi-scale convolution, attention mechanism and dynamic feature regulation, the processed deep layer scale feature map and other multi-scale feature maps are input into the adaptive hierarchical feature path aggregation network, each scale feature map of each input is fused from top to bottom and from low to up, and the cross-scale fusion result and each scale feature map of each input are fused and optimized by using the adaptive hierarchical feature fusion unit to obtain the optimized fusion feature map, and the detection head is used for target detection according to the optimized fusion feature map to obtain a detection result. The method can effectively improve the target detection efficiency and accuracy of the infrared image. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1A flowchart of an infrared image target detection method in an embodiment is shown in FIG. 1. Figure 2 A schematic diagram of the overall framework of a target detection network in an embodiment is shown in FIG. 2. Figure 3 A schematic diagram of the structure of an adaptive weight division unit in an embodiment is shown in FIG. 3. Figure 4 A schematic diagram of the structure of a polarization dynamic attention fusion unit in an embodiment is shown in FIG. 4. Figure 5 A schematic diagram of the structure of an adaptive hierarchical feature path aggregation network in an embodiment is shown in FIG. 5. Figure 6 A schematic diagram of the structure of an aggregation block in an embodiment is shown in FIG. 6. Figure 7 A schematic diagram of the structure of an adaptive hierarchical feature fusion unit in an embodiment is shown in FIG. 7. Figure 8 A schematic diagram of the structure of a local-global attention block in an embodiment is shown in FIG. 8. Figure 9 A visualization diagram of the extracted features of the feature extraction network in the present method and the RT-DETR network in an experiment is shown in FIG. 9. Figure 10 A comparison diagram of the extracted features of the RT-DETR network and the feature enhancement using the polarization dynamic attention fusion unit in the present method in three different infrared images in an experiment is shown in FIG. 10. Figure 11 A comparison diagram of the target detection results of the YOLOv8m network, the RT-DETR network, and the model in the present method in four tests in an experiment is shown in FIG. 11. Figure 12 A result diagram of an ablation experiment in an experiment is shown in FIG. 12. Figure 13 A block diagram of the structure of an infrared image target detection device in an embodiment is shown in FIG. 13. Figure 14 An internal structure diagram of a computer device in an embodiment is shown in FIG. 14. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0019] In view of the problems existing in the prior art of using a neural network to detect targets in an infrared image, as shown in FIG. 1, an infrared image target detection method is provided in the present embodiment, which specifically includes the following steps: Figure 1 ​Step S100, an infrared image to be subjected to target detection is acquired.

[0020] Step S110, the infrared image is input into a feature extraction network for feature extraction to obtain a multi-scale feature map, wherein an adaptive weight division unit is used in the feature extraction network for feature extraction, and the adaptive weight division unit realizes real-time selection and adaptive adjustment of a convolution kernel through a dynamic convolution kernel weighting mechanism.

[0021] Step S120, a deep layer scale feature map in the multi-scale feature map is processed using a polarized dynamic attention fusion unit to obtain a processed deep layer scale feature map, and the polarized dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep layer scale feature map through multi-scale convolution, attention mechanism and dynamic feature regulation.

[0022] Step S130, the processed deep layer scale feature map and other multi-scale feature maps are input into an adaptive hierarchical feature path aggregation network, each scale feature map of each input is cross-scale fused from top to bottom and from low to high, and the cross-scale fused result is fused and optimized with each scale feature map of the input using an adaptive hierarchical feature fusion unit to obtain an optimized fused feature map.

[0023] Step S140, a detection head is used to perform target detection according to the optimized fused feature map to obtain a detection result.

[0024] In step S100, the infrared image to be subjected to target detection can be a thermal image converted from infrared radiation emitted by a scene itself. Since different parts of an object have different radiation characteristics, the infrared image can distinguish different parts of the scene.

[0025] In step S110, considering that in a traditional convolutional neural network (CNN) architecture, a convolution kernel is usually designed as a fixed and static entity, the convolution kernel increasingly exposes inherent limitations when facing complex, multi-scale and highly variable input data. Specifically, the singularity of kernel selection limits the ability of the network to fully adapt to different scale features, which is a key defect, and is particularly evident when processing images with significant scale changes, in which a single kernel cannot capture both fine-grained details and global context information. In addition, strict kernel design leads to low computational efficiency and suboptimal parameter utilization, because excessive computation and parameters not only increase memory consumption, but also cause processing bottlenecks, which are particularly critical in real-time applications and large-scale data set scenarios.

[0026] In this embodiment, in order to solve these problems while meeting the real-time detection requirements, a new architecture module, the adaptive weight division unit (AWD-Block), is proposed. In this unit, a dynamic convolution kernel weighting mechanism is innovatively integrated, enabling real-time selection and adaptive adjustment of the convolution kernel. By dynamically allocating kernel weights according to the immediate input feature requirements, efficient capture of multi-scale and heterogeneous feature information is achieved, establishing a robust feature foundation for real-time detection tasks. In addition, the network employs a multi-scale convolution kernel fusion strategy, enabling cross-scale feature extraction and information sharing, greatly enhancing the model's adaptability to complex scenes and ensuring consistent detection accuracy even under challenging real-world conditions. Through this carefully designed architecture, the adaptive weight division network significantly improves scene adaptability while achieving significant computational efficiency, thereby robustly ensuring real-time detection performance. Its unit structure is as shown in FIG. 1. Figure 3

[0027] In this embodiment, in the adaptive weight division unit: the input image is split into two parts and input into two parallel adaptive weight division convolution branches. In the adaptive weight division convolution branch, adaptive average pooling, dynamic kernel weight generation, and weighted convolution are used for differential feature extraction of the input data. Then, through concatenation and fusion, the feature expression is enhanced. The data output by the two parallel adaptive weight division convolution branches is then concatenated and passed through a convolution layer to obtain a feature map of a certain scale.

[0028] Specifically, the input data is first split into two parts through the "Split" operation and sent to two parallel adaptive weight division convolution branches. The results of the two adaptive weight division convolution branches are concatenated through the "Concat" operation, and then a "Conv" (convolution) is performed to complete the adaptive weight division network process, realize the depth processing and feature fusion of the input data, and be used for subsequent feature extraction and other tasks of the network.

[0029] Further, the adaptive weight division convolution branch uses square convolution kernel stripes along the horizontal and vertical directions to capture features and scales in different directions. Using a dynamic weight generation mechanism, it adaptively fuses the results from these convolution branches. The channel grouping strategy combines multi-scale convolution kernels with 1x1 convolution to enable cross-channel information interaction. Finally, the dynamic weight mixing module integrates a multi-path residual structure, combining dynamic scaling parameters and random depth regularization. This design maintains computational efficiency and training stability while significantly enhancing the model's feature extraction capability for complex textures and multi-scale targets.

[0030] ​Specifically, the adaptive weight-based convolutional branching, by combining square convolutional kernels with horizontal and vertical striped convolutional kernels, effectively captures features at different directions and scales. The network employs a dynamic weight generation mechanism that adaptively fuses the outputs of each convolutional branch. Its channel grouping strategy combines multi-scale convolutional kernels with 1×1 convolutions, promoting cross-channel information interaction. The dynamic weight fusion module integrates a multi-path residual structure, incorporating dynamic scaling parameters and stochastic depth regularization. This design significantly enhances the model's ability to extract features from complex textures and multi-scale targets, while ensuring computational efficiency and training stability. This approach provides a new paradigm for building optimized deep learning models, particularly suitable for tasks with variability and complexity in input data.

[0031] In this embodiment, a feature extraction unit (AWD-Net) for multi-scale feature extraction is constructed based on the aforementioned adaptive weight partitioning unit. The feature extraction network includes convolutional layers and multiple feature extraction blocks, each consisting of a convolutional layer and the adaptive weight partitioning unit. The input image is sequentially input to multiple feature extraction blocks after passing through a convolutional layer, and the adaptive weight partitioning unit in each feature extraction block outputs feature maps of different scales.

[0032] In one embodiment, such as Figure 2 As shown, in the feature extraction network, the infrared image passes through a convolutional layer, and then sequentially through four feature extraction blocks consisting of a convolutional layer and an adaptive weight partitioning unit. Starting from the second feature extraction block, multi-scale feature maps are output from the weight partitioning unit, which are arranged from shallow to deep as shallow scale feature map S3, medium scale feature map S4, and deep scale feature map S5. These multi-scale feature maps are then input into the multi-scale feature fusion network for feature fusion.

[0033] In step S120, before the multi-scale feature fusion of the deep-scale feature map S5, a polarization dynamic attention fusion unit (PDAF) is also used for processing, which overcomes the calculation bottleneck of the traditional transformer through the polarization linear attention mechanism, adopts a 20x20 grid-based partition strategy to flatten the spatial dimension into sequence data for linear attention calculation. This method reduces the complexity from O(n2) to O(n) while maintaining global context awareness, improving efficiency. While preserving the lightweight design of the model, this work lays a solid foundation for real-time detection capability. The SE Feedforward Network in it implements double-path feature processing: the spatial branch uses a down-sampling-convolution-up-sampling pipeline to extract multi-scale context information, which is then fused with the cross-scale features of the main branch through a depthwise separable convolution. After dynamic parameter modulation, the output is adaptively weighted features. The multi-cognitive visual adapter uses a 64-channel bottleneck layer for non-linear feature conversion, and through a double-path dynamic mixing mechanism, it realizes channel attention similar to squeezing and exciting, with a 30% reduction in computational overhead. DyT replaces the traditional layer normalization, retaining numerical stability while enhancing non-linear modeling capabilities, resulting in a 40% improvement in model convergence speed. These synergistic technologies form a highly efficient feature modeling framework that performs outstandingly in infrared image detection scenarios. The detailed architecture of the polarization dynamic attention fusion unit is shown in Figure 4

[0034] In this embodiment, the polarization dynamic attention fusion unit takes the deep-scale feature map and the corresponding position encoding as double-path input, first uses the Polo linear attention mechanism for long-range association, uses the ADD block to realize feature accumulation, uses the DyT block to dynamically regulate the features, uses the SE Feedforward Network to focus on key channel features, and finally uses the Mona block for integration.

[0035] As shown in Figure 4 , after the deep-scale feature map and the corresponding position encoding are input into the polarization dynamic attention fusion unit, they are first processed using the Polo linear attention mechanism, then input into the first group of ADD blocks and DyT blocks, and their output is processed through the first Mona block and the SE Feedforward Network, and then through the second group of ADD blocks and DyT blocks, and finally integrated by the second Mona block.

[0036] ​In the present embodiment, in the Mona block: the input data is first normalized, then dimensionally reduced by the Down Projection layer, the data after dimension reduction is divided into multiple paths, and features of corresponding scales are extracted by convolution layers with different depths respectively, the features of multiple scales are fused by the Average layer, and then residual fusion is performed with the data before convolution, and the data obtained after fusion is further subjected to a convolution layer, a GeLU activation layer and an Up Projection layer for dimension increase. Among them, after the convolution layer, residual fusion of data is also performed, and then the data after dimension increase is subjected to residual fusion with the input data, to complete multi-scale feature depth integration and enhancement.

[0037] The Mona block is as shown in the left side of the figure. Figure 4 As shown in the left side of the figure, the data after dimension reduction is subjected to multi-scale feature extraction by parallel 3x3, 5x5 and 7x7 depth separable convolution kernels, so as to effectively capture multi-scale spatial features from visual signals. The scaling normalization layer (Down Projection) dynamically adjusts the input feature distribution to optimize the performance of the visualization task. The 1x1 convolution reduces and then restores the number of channels. The skip connection promotes functional flow and alleviates the problem of gradient disappearance. During operation, the Mona block first optimizes the input feature distribution, then extracts and fuses features by multi-scale convolution, and finally outputs enhanced features by dimension expansion and residual connection.

[0038] Specifically, in the Mona block, the input data is first processed by Layer Norm (corresponding to the scaling normalization layer, which dynamically adjusts the input feature distribution to optimize the performance of the visualization task), enters Down Projection (similar to the dimension reduction layer, which reduces the number of channels by 1x1 convolution), and then enters the multi-cognitive visual filter library, which captures multi-scale spatial features from visual signals by parallel 3x3, 5x5 and 7x7 depth separable convolution (DWconv), and then goes through Average (average pooling, which captures multi-scale spatial features and reduces computational load), Concat (fuses multi-scale features), GELU (activation function), Up Projection (dimension expansion), and finally through the skip connection (represented by the body of in the figure, which promotes functional flow and alleviates the problem of gradient disappearance), to complete the whole process of input feature from optimized distribution, to multi-scale convolution extraction and fusion, to dimension expansion and residual connection output enhanced features.

[0039] Further, the data flow in the Mona block can be represented as:

[0040]

[0041]

[0042]

[0043] In formulas (1) to (4), given the input tensor Layer normalization is first applied to distribution calibration, where and The channel mean and standard deviation represent the input features. Learning parameters. (Scaling) and (Shift) and numerical stability constant Used for feature normalization. Then, preprocessing is performed using parallel multi-scale convolutional kernels for feature extraction, where, This represents a depthwise separable convolution operation with a kernel size of k×k. Multi-scale features are obtained by fusion using average pooling. Then, channel size operations are performed using 1×1 convolution to compress the channel. and recovery relative to the original number of channels C This method effectively solves the computational bottleneck of full parameter fine-tuning, significantly reduces computational overhead and memory consumption, and overcomes the performance limitations of traditional methods in complex vision tasks.

[0044] In this embodiment, the Enhanced Channel Attention Block (SE Feedforward Network) is connected after the Mona block. In the Enhanced Channel Attention Block, a dual-branch processing method is used to process the features obtained after passing through the Mona block and the features that have not passed through the Mona block. The features obtained after passing through the Mona block are transformed through a conventional feature transformation through one branch, while the features that have not passed through the Mona block are mined through multi-scale convolution and upsampling through another branch. The output features of the two branches are concatenated and fused, and intelligent weighting is performed using a gating mechanism to achieve refined integration and enhancement of features.

[0045] Furthermore, such as Figure 4 As shown in the right panel, the Enhanced Channel Attention Block (SE Feedforward Network) architecture consists of the input... (Features not derived from Mona blocks) and The features (obtained after passing through the Mona block) consist of two feature capture branches. To enhance spatial awareness, Average pooling is performed to expand the receptive field, followed by two sequential convolutional layers with normalized ReLU blocks to capture spatial context. Upsampling is then used to generate spatial awareness metrics. In gating mechanisms, Branches are divided into and ,in Before generating the gate signal, the convolution and GELU activation are combined. The gate signal is multiplied element-wise with to modulate its spatial sensitivity.

[0046] Specifically, the channel attention block architecture is strengthened to input the module and Two feature capture branches are constructed to enhance spatial awareness, wherein, in the branch, the first average pooling (Average) extended receptive field is performed, then the spatial context is captured through two "Conv 3x3-norm- ReLU" blocks, and then up-sampling (Upsample) is performed to generate a spatial awareness index In the branch, the is divided into and and , wherein, After the convolution (Conv 1x1) and GELU activation, the gate signal (gate) is generated in combination with , which is multiplied element-wise with to modulate its spatial sensitivity, and finally the spatial enhancement unit output is obtained by multiplying the two, realizing the enhancement and fine modulation of the feature spatial awareness.

[0047] Further, in the channel attention block (SE Feedforward Network), the process is represented as:

[0048]

[0049]

[0050]

[0051]

[0052] In formulas (5) to (9), and represent 1x1 convolution layers, and represent 3x3 depth separable convolution, correspond to layer normalization, represent the convolution layer normalization ReLU block, represent the up-sampling operation.

[0053] ​In this embodiment, the Polo Linear Attention mechanism in the Polar Dynamic Attention Fusion Unit solves the limitations of traditional linear attention mechanisms that ignore negative information. This is achieved by explicitly modeling the positive and negative interactions between the query-key pairs Q-K. Specifically, the mechanism decomposes the query vector Q and the key vector K into positive components , and negative components 、 . Then the attention weight is calculated as:

[0054] This design captures both the similarity between components of the same type and the interaction between components of different types, preventing the loss of negative information.

[0055] Further, for the block, it replaces the traditional normalization layer with a dynamic hyperbolic tangent function, defined as:

[0056] In the above formula, represents a learnable scaling parameter, while and represent channel affine transformation parameters. By appropriate scaling coefficients, and using the nonlinearity of to compress extreme values, the mean square deviation adjustment behavior of the normalization layer is simulated without explicit statistical calculation.

[0057] In deep learning, especially in computer vision tasks, multi-scale feature fusion and attention mechanisms have been widely verified, significantly enhancing the model's perception and understanding of complex scenes. However, a key challenge faced by current multi-source feature fusion methods is the difficulty in effectively coordinating and integrating local details and global semantic context. To solve this problem, in this method, the Adaptive Hierarchical Feature Path Aggregation Network (AHF-PAN) is introduced, and its core is the Adaptive Hierarchical Feature Fusion Unit (AHFF-Block).

[0058] In the embodiment, the design concept of the adaptive hierarchical feature fusion unit is to realize adaptive fusion of local and global features through hierarchical attention mechanism, improve the robustness and task adaptability of feature representation, and the key innovation is to integrate local and global attention branches through a local-global attention block and a hierarchical feature processing path for fine integration. The method preserves the spatial details of the input features, and uses an adaptive weighting mechanism to enhance the semantic expression of the global context, accelerate feature focusing and enhance fusion fluency. Such optimization provides a basic framework for real-time infrared image detection tasks. The detailed architecture of the adaptive hierarchical feature path aggregation network is as shown in Figure 5

[0059] As shown in Figure 5 In the adaptive hierarchical feature path aggregation network, the core aggregation block realizes the integration of the two parts of top-down feature fusion and bottom-up feature fusion. Initially, the multi-scale feature maps are aligned through operations such as upsampling or downsampling to achieve the same spatial resolution. Subsequently, the multi-scale feature maps are connected along the channel dimension and fused through a 1x1 convolution layer to finally generate multi-scale intermediate fusion feature maps.

[0060] Specifically, the processed deep layer scale feature map F5 generates a deep layer intermediate fusion feature map A5 through a convolution layer. After upsampling, the deep layer intermediate fusion feature map A5 is aggregated with the middle layer scale feature map S4 in the aggregation block, and a middle layer intermediate fusion feature map A4 is generated through a convolution layer. After downsampling, the middle layer intermediate fusion feature map A4 is aggregated with the shallow layer scale feature map S3 in the aggregation block to generate a shallow layer intermediate fusion feature map A3, thereby completing the top-down cross-feature fusion.

[0061] Further, the shallow layer intermediate fusion feature map A3, i.e., the shallow layer fusion feature map B3, is aggregated with the middle layer intermediate fusion feature map A4 in the aggregation block after downsampling to generate a middle layer fusion feature map B3. After downsampling, the middle layer fusion feature map B3 is aggregated with the deep layer intermediate fusion feature map A5 in the aggregation block to obtain a deep layer fusion feature map B5.

[0062] In the embodiment, the aggregation block enhances the feature representation capability by merging detail and semantic information, thereby improving the detection performance of the model across multiple scales. In addition, it promotes information exchange between different scale features, providing more abundant and effective features for subsequent detection tasks. The detailed structure of the aggregation block is as shown in Figure 6

[0063] ​​Specifically, in the aggregation block, the input multiple features are first spliced by the "Concat" operation to integrate the feature information of different sources or different dimensions. Then, the spliced features are input into two parallel branches for processing. In one branch, the spliced features are processed by convolution (Conv1x1) to quickly adjust the channel dimension and extract basic features. In the other branch, after the initial processing of the spliced features by convolution (Conv1x1), the features are sent to the reparameterization block (RepBlock) for feature enhancement processing. With the help of the reparameterization technique, the feature representation is optimized in the training and inference stages, and the feature extraction capability is enhanced. Then, the results of the two parallel branches are fused by element-wise addition, and finally, the spatial dimension features are flattened by the "Flatten" operation to compress the spatial dimension features into one-dimensional sequences, which facilitates the subsequent processing of tasks such as classification and regression, realizes feature aggregation and dimension conversion, and provides effective input for the subsequent process of the network.

[0064] In this embodiment, the polarization dynamic attention fusion unit can also be regarded as part of the adaptive hierarchical feature path aggregation network. The data flow in the adaptive hierarchical feature path aggregation network can be simply represented as:

[0065]

[0066]

[0067] In this embodiment, in the adaptive hierarchical feature path aggregation network, the shallow fusion feature map B3, the middle fusion feature map B3, and the deep fusion feature map B5 are obtained. The adaptive hierarchical feature fusion unit (AHFF) optimizes the original feature maps corresponding to each fusion feature map, as shown in Figure 5 .

[0068] In this embodiment, in the adaptive hierarchical feature fusion unit, the two input features are first reduced in dimensionality by a convolutional layer to obtain a first reduced-dimensionality feature and a second reduced-dimensionality feature, respectively, unifying the feature dimensions while reducing the complexity of subsequent calculations. Then, the first and second reduced-dimensionality features are summed and further fused by a convolutional layer to generate a baseline feature. Next, the first and second reduced-dimensionality features are optimized by a local-global attention block to obtain a first optimized feature and a second optimized feature. The baseline feature, the first optimized feature, and the second optimized feature are fused using a "Concat" operation, and then sequentially passed through a composite block consisting of a convolutional layer, a reparameterized convolutional layer, and convolutional layers (Conv1×1, RepConv3×3, Conv1×1). The initial convolutional layer is mainly used to reduce the dimensionality of the input features and computational overhead, employing reparameterized convolutions for feature recombination. This method significantly improves parameter utilization efficiency and can better integrate features from multiple branches.

[0069] like Figure 7 He Ru Figure 8 As shown, the local-global attention block includes two parallel attention branches with different parameter settings. Here, the parameter, denoted by p, is a key parameter for the patch operation and defines the granularity of local feature extraction. A smaller p value results in a smaller patch size and finer feature extraction granularity, which is beneficial for capturing local details such as edges and textures of small objects. However, excessively small p values ​​may lead to increased computational complexity and the introduction of potential noise. Conversely, larger p values ​​result in larger patch sizes and a wider receptive field, enabling the capture of more contextual information.

[0070] In this embodiment, p=2 and p=4 represent local attention and global attention, respectively. The local attention branch focuses on extracting local detail features from the infrared image. Specifically, this branch first partitions the input feature mapping mechanism, and during the weighting process, a feature selection method is used to select task-relevant features from local window regions and channels. Subsequently, attention weights between patches are calculated to highlight key local features. This process can be mathematically expressed as follows:

[0071] In formula (4), This represents the input features that are fed into the local attention branch. This indicates a patch operation. This represents the linear projection in the local attention branch. This represents the attention mechanism within the local attention branch.

[0072] In the present embodiment, the global attention branch is responsible for capturing global context information from the infrared image. The branch initially performs global average pooling on the input feature map to obtain a global feature vector. In order to enhance the model's perception of the overall structure of the target, the global branch further expands its receptive field through a shift operation. Finally, the global feature is combined with the local feature to improve the distinguishability of the target and the background. This process can be expressed in mathematical formula as:

[0073] In formula (5), represents the input feature to the global attention branch, represents the shift operation, represents the global average pooling, represents the feature concatenation, represents the attention mechanism in the global attention branch.

[0074] Specifically, in the weighting process, a feature selection method is used to select task-related features from tokens and channels. That is, let represent the dimension, and the weighted result is represented as where represents the i-th output token. Subsequently, a linear transformation is applied to p to select the channels of each token, followed by reshaping and interpolation operations. This finally generates and respectively through the localized and globalized feature refinement paths. In step S140, the final fusion feature output by the adaptive hierarchical feature path aggregation network is input into the detection head. The final fusion feature is first screened by the minimum uncertainty query selection unit, and then sent to the decoder and the detection head for processing to obtain the target detection result.

[0075] In this paper, the effectiveness of the proposed neural network architecture is also demonstrated through experiments. In the experiments, the infrared image detection dataset used is from the Roboflow platform, which serves as a reliable benchmark for evaluating and enhancing object detection capabilities in complex environments. The dataset includes a wide range of scenarios, including outdoor, street, and urban environments, captured through drone aerial photography and fixed camera surveillance systems. It consists of 4,918 infrared images, which are divided into training, validation, and test sets in a 7:2:1 ratio. The dataset has 16 object classes, including "bicycle," "bus," "car," "deer," "dog," "fire hydrant," "light," "motor," "other vehicle," "person," "scooter," "sign," "skateboard," "stroller," "train," and "truck." Diversity is ensured by considering factors such as perspective, lighting conditions, weather changes, and object-specific factors such as size, shape, and occlusions. This results in significant variations in the dataset's scale, pose, and object density, providing a rich training sample for the model. All images are preprocessed to a uniform resolution of 640x640 pixels and meticulously annotated with object class, bounding box coordinates, and size information, providing strong support for model training and evaluation.

[0076] In the experiments, a visual comparison of the features was first conducted. In deep learning models, the backbone network primarily serves as a feature extraction module, processing the global structure of the input data. Through a series of complex operations, including convolution and pooling, the backbone network captures key information and encodes it into discriminative feature representations. However, feature maps are an intuitive visualization tool that can effectively demonstrate the feature extraction capabilities of the backbone network. In these maps, yellow regions highlight the network's activation of target objects and its fine learning of relevant features, indicating successful capture of key object characteristics. In contrast, blue regions reflect the network's effective suppression of background information, which helps to mitigate interference from irrelevant context and improve the model's accuracy and robustness. Detailed visualization results are shown in Figure 9 .

[0077] As shown in Figure 9 , in the study of the RT-DETR model, ResNet-18 was used as its feature extraction backbone network in the experiments. However, as the visualized feature maps demonstrate, this model is ineffective in feature extraction of target objects, with a significant portion of feature learning focused on background regions and non-target elements. This phenomenon has a negative impact on both detection accuracy and computational efficiency. In contrast, the AWD-Net model proposed in this method demonstrates superior feature extraction capabilities, accurately learning target contour features while maintaining high sensitivity to point and line features. Notably, it achieves effective background suppression by minimizing interference from irrelevant contextual information, significantly improving the reliability and accuracy of object detection.

[0078] Next, the heat map comparison experiment of the polarization dynamic attention fusion unit (PDAF) is carried out. In order to provide more intuitive and clear visualization of the attention area processed by the polarization dynamic attention fusion unit (PDAF) during feature extraction, heat maps are used as effective analysis tools in the experiment. These heat maps intuitively represent the regional attention distribution of the model through color intensity gradients. Specifically, darker tones indicate that the model allocates higher attention weights to the corresponding region during feature processing, reflecting greater emphasis, while lighter tones indicate relatively lower importance. This visualization method helps to directly observe the focal regions and attention dynamics of the model, and the comparison results are shown in Figure 10 .

[0079] As shown in Figure 10 , in order to comprehensively evaluate the performance of different models in the infrared object detection task, three representative test images are selected in the experiment, covering different scenes and target features, thus thoroughly verifying the ability of the model in complex environments.

[0080] Specifically, in the image 1 test scene featuring small target objects, the experimental results show that the RT-DETR model exhibits significant bias in feature extraction and attention focusing, often misdirecting attention to non-target objects. These distracted attention may have visual similarities in color, shape or texture with the actual target, leading to model errors. However, when tested using the polarization dynamic attention fusion unit (PDAF), the visualized heat map clearly indicates that the red region is accurately aligned with the target object. This observation confirms that the polarization dynamic attention fusion unit (PDAF) effectively guides the model to focus on the true target, greatly improving the detection accuracy of small pixel objects. In the dense object scene of image 2, the polarization dynamic attention fusion unit (PDAF) demonstrates excellent attention allocation ability, accurately allocating attention between multiple target objects and successfully avoiding confusion and false detection caused by object crowding. For the complex background test of image 3, the complex background elements are intertwined with the target objects, causing significant interference, and the PDAF enhanced model maintains its outstanding performance by accurately capturing the target in the visual clutter. This further verifies the excellent attention focusing ability of the module in challenging environments.

[0081] In this experiment, a series of rigorous comparative experiments were designed when evaluating the overall model. For the selection of evaluation indicators, a holistic and multi-dimensional approach was adopted. Specifically, focus was placed on the model parameter indicator, which effectively measures the complexity and resource consumption of the model during operation. The GFLOPS indicator was used to evaluate the computational efficiency of the model, directly reflecting its data processing speed. The recall indicator clearly demonstrates the model's ability to identify positive class instances, indicating how many actual positive samples were correctly identified. The precision indicator emphasizes the importance of ensuring the accuracy of the model's prediction results while reducing the likelihood of false judgments. Additionally, the mAP@0.5 and mAP@0.5:0.95 indicators were used, which serve as average precision indicators at different IoU thresholds. These indicators provide a comprehensive evaluation of the model's performance under various detection difficulties. For the selection of comparative models, widely recognized benchmark models were chosen, including DETR, deformable DETR, Faster R-CNN, SSD, and the YOLO series, to facilitate in-depth comparative analysis. Detailed comparison results are shown in Table 1.

[0082] Table 1 Performance comparison between different models

[0083] As shown in Table 1, the parameter count and GFLOP of AD-DETR are significantly reduced compared to models with larger parameter counts and GFLOP, such as SSD, Faster R-CNN, and Deformable DETR. At the same time, it achieves a performance improvement of 9.3%, 11.0%, and 11.9% in the mAP@0.5 indicator, respectively. To comprehensively evaluate the performance of AD-DETR, YOLO series models with comparable parameter counts and GFLOP were selected for comparison, including YOLOv5m, YOLOv6s, YOLOv8m, and YOLOv10m. This comparison result shows that the precision of AD-DETR is the highest, reaching 77.6%. In the mAP@0.5 indicator, AD-DETR demonstrates superior performance, with an increase of 5.2%, 7.5%, 6.0%, and 6.6% over the aforementioned models, respectively. In addition, in the mAP@0.5:0.95 indicator, AD-DETR maintains a clear advantage, exceeding these models by 7.9%, 8.7%, 6.1%, and 5.4%, respectively. These findings indicate that AD-DETR can more accurately capture the detailed features of targets, effectively reducing false positives and false negatives. Furthermore, compared to the DETR and RT-DETR models, the precision of AD-DETR increases by 9.2% and 2.9%, respectively, the mAP@0.5 increases by 9.8% and 4.9%, and the mAP@0.5:0.95 increases by 10.7% and 4.5%. This further emphasizes the stability and robustness of AD-DETR under different detection difficulties.

[0084] Table 2. Detection results of AD-DETR model on infrared image dataset

[0085] As shown in Table 2, the effectiveness of the target detection model in various infrared object categories is comprehensively demonstrated. Overall, the model shows excellent performance in all indicators, with a precision of 77.6%, a recall rate of 71.5%, and mAP@0.5 and mAP@0.5:0.95 scores of 68.4% and 43.8%, respectively. These results powerfully verify its significant efficacy in high-precision infrared object detection. For transportation categories such as "car", "bus", "motor", "train", and "truck", the model shows consistent high detection accuracy. Specifically, the precision of "car" reaches 80.9%, and mAP@0.5 climbs to 80.9%. For "bus", "motor", "train", and "truck", the model achieves precision of 70.5%, 86.5%, 86%, and 64.3%, respectively, and mAP@0.5 scores of 55.2%, 84.4%, 81.2%, and 42.7%, respectively. This indicates that the model can stably and accurately detect different types and sizes of transportation targets. For common objects, including "bicycle", "dog", "person", and "stroller", the model also shows good performance in detecting these daily targets. For smaller targets such as "scooter" and "skateboard", which usually occupy fewer pixels in infrared images and pose greater detection challenges, the model achieves perfect precision scores of 100% in both categories, with mAP@0.5 scores of 88.7% and 89.8%, respectively. The model shows excellent detection performance for other categories, including "fire hydrant", "light", "sign", "deer", and "other vehicle".

[0086] Further, visual comparison of detection results is conducted, and in the experiment, four representative images are selected from different scenes to visually compare the target detection performance. These scenes cover various aspects of infrared image object detection, including small targets, dense occlusion, infrared radiation interference, and low resolution challenges. To comprehensively evaluate the performance of different models in these scenes, two advanced target detection models, YOLOv8m and RT-DETR, are selected for comparison.

[0087] As Figure 11As shown, the detection performance of these models in selected scenarios is compared. Through intuitive visual demonstrations, we can clearly observe the differences in the ability of various models to capture target details, distinguish targets from backgrounds, and manage complex scenes. This visual comparison not only enhances the understanding of the strengths and weaknesses of each model, but also provides substantial support for subsequent performance analysis and optimization. Test 1 focuses on the detection performance of small targets by the model. In this test scenario, the target objects are very small, and the ability of the model to capture details poses a major challenge. After careful observation, it is noted that other models exhibit varying degrees of false detection, with some small targets being completely ignored. In contrast, the AD-DETR model performs exceptionally well, successfully detecting all small targets in the field without any omissions, while also greatly improving the confidence of the detection results. This indicates that the AD-DETR model not only accurately identifies small targets, but also provides more reliable evidence for judgment.

[0088] Test 2 mainly demonstrates the detection performance of the model under dense occlusion conditions. In the test image, the vehicles and pedestrians on the left are intertwined and tightly blocked, greatly increasing the difficulty of detection. The results show that the YOLOv8 model fails to accurately identify all targets in such complex occlusions, missing several occluded targets. In contrast, the AD-DETR model performs well, successfully detecting all targets. In addition, compared with the RT-DETR model, the AD-DETR model performs better in terms of detection accuracy and integrity, thereby significantly improving the overall detection effect.

[0089] Test 3 simulates an external infrared radiation interference to evaluate the detection ability of the model under complex lighting conditions. In the right side of the test image, the pedestrian is interfered by infrared radiation, partially blurring the target features. Comparison shows that only the AD-DETR model successfully identifies this pedestrian, while other models fail to detect it. This finding highlights the superior detection ability of the AD-DETR model in the presence of external interference, maintaining high detection accuracy in complex environments.

[0090] To further verify the performance of the model under extreme conditions, low-pixel images are selected for Test 4 to simulate the detection effect when image quality deteriorates due to device damage or adverse weather. During the test, it is observed that the other two models fail to accurately identify targets to varying degrees, with some models exhibiting missed detections and others exhibiting false detections. In contrast, the AD-DETR model exhibits excellent recognition ability, accurately identifying targets despite the low image resolution. This further confirms the stability and reliability of the model in various complex scenarios.

[0091] Further, ablation experiments are conducted. To intuitively demonstrate the impact of the proposed improvements on the overall detection performance of the model, six groups of ablation experiments are conducted. Each enhancement is integrated into the network step by step, and the contribution of each part is evaluated. Table 3 systematically summarizes the results of these experiments.

[0092] Table 3 Ablation Experiments

[0093] As shown in Table 3, a comparative analysis of the performance of various model configurations in the infrared image object recognition task is performed. The baseline model serves as a reference point, illustrating the initial computational load, inference speed, parameter count, and detection accuracy. The integration of the AWD-Net component significantly improves detection performance, with F1 score and mAP@0.5 increasing by 1.6% and 3.7%, respectively. At the same time, the computational load is significantly reduced, and the inference speed is significantly accelerated, indicating that AWD-Net effectively improves feature extraction efficiency and model inference speed. In addition, the combination of PDAF and AHF-PAN components, although resulting in a slight increase in computational load, enhances the multi-scale detection capability of the model, leading to continuous improvement in performance indicators. Notably, in Model 5, the integration of AWD-Net and PDAF modules results in a 2.3% increase in mAP@0.5:0.95 compared to using AWD-Net alone, highlighting the synergistic advantages of multi-component collaboration. Finally, the optimized model integrating all enhancement strategies achieves the highest performance, with an F1 score of 74.4% and an mAP@0.5:0.95 of 43.8%. Compared with the baseline model, these numbers improve by 3.7% and 4.5%, respectively. These results not only verify the effectiveness of each enhancement module but also emphasize the significant performance benefits brought by their collaborative operation, providing an efficient and accurate solution for infrared image object recognition.

[0094] To intuitively evaluate the detection performance of each module, their impact on the target detection task is also visually demonstrated in the experiments. By systematically comparing the detection performance of different modules, their respective advantages and limitations can be clearly divided, providing a foundation for model optimization. Detailed visualization results are shown in Figure 12 .

[0095] In the visualization demonstration section, representative challenging infrared images are selected to intuitively demonstrate the detection effect. The characteristics of this image are that the target size is relatively small, and there is obvious occlusion between targets. It contains four types of objects to be detected, which not only tests the robustness of the model but also puts high demands on its ability to identify various target types.

[0096] As Figure 12(a) The detection results of the baseline model show low overall confidence and missed detections. This indicates that it is insufficient in identifying small targets and also fails to effectively manage complex and varying real-world scenarios. Figure 12 (b) and (c) illustrate the detection results after integrating the PDAF and AWD-Net modules sequentially. It is observed that the model successfully detects some previously missed targets and shows an improvement in overall confidence. However, there are still some instances where objects are not detected, indicating the need for further improvement. Finally, as shown in Figure 12 (d) shows that the integration of the AHF-PAN module greatly improves the detection performance due to its superior feature fusion capabilities. This enables the model to accurately identify objects under complex conditions, not only demonstrating improvements in small target detection but also enhancing the ability to manage various complex situations such as occlusions, overlaps, and background disturbances.

[0097] In the above infrared image target detection method, by introducing the adaptive weight division network, the dynamic selection and adaptive adjustment of the convolution kernel are realized, effectively enhancing the model's ability to extract features from infrared images. In addition, the adaptive hierarchical feature path aggregation network utilizes a hierarchical attention mechanism to promote the adaptive fusion of local and global features, thereby enhancing the model's resilience in complex backgrounds. Furthermore, the polar dynamic attention fusion unit overcomes the computational bottleneck of traditional transformers, enabling cross-scale feature fusion and significantly improving the model's detection accuracy for small targets.

[0098] At the same time, experiments on infrared image datasets show that AD-DETR exhibits high precision and strong robustness, effectively meeting the needs of infrared image target detection in complex environments. This model is particularly suitable for scenarios with low signal-to-noise ratios, complex background disturbances, and significant target size variations. These scenarios include obstacle detection in autonomous driving environments, anomaly identification in urban surveillance systems, and personnel or target positioning in forest search and rescue operations. Due to its fast detection speed and low overall complexity, AD-DETR can be deployed on resource-constrained edge computing devices or embedded systems that require real-time responses.

[0099] It should be understood that, although Figure 1 The steps in the flowchart of FIG. 1 1 are not necessarily performed in the order indicated by the arrows. Unless explicitly stated otherwise herein, the steps of the flowchart can be performed in any order. Moreover, Figure 1At least one of the steps in the method can comprise a plurality of sub-steps or a plurality of stages, which sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the order of the sub-steps or stages is not necessarily sequential, but can be performed alternately or in rotation with other steps or sub-steps or stages of other steps.

[0100] In one embodiment, as shown in Figure 13 An infrared image target detection method and device are provided, comprising: an image acquisition module 200, a multi-scale feature extraction module 210, a deep scale feature map optimization module 220, a multi-scale feature fusion module 230, and a target detection module 240, wherein: The image acquisition module 200 is configured to acquire an infrared image to be subjected to target detection. The multi-scale feature extraction module 210 is configured to input the infrared image into a feature extraction network to extract features, thereby obtaining a multi-scale feature map. The deep scale feature map optimization module 220 is configured to process a deep scale feature map in the multi-scale feature map using a polarized dynamic attention fusion unit, thereby obtaining a processed deep scale feature map. The multi-scale feature fusion module 230 is configured to input the processed deep scale feature map and other multi-scale feature maps into an adaptive hierarchical feature path aggregation network, to perform top-down and low-to-up cross-scale fusion on each scale feature map, and to fuse and optimize the results of the cross-scale fusion and the input scale feature maps using an adaptive hierarchical feature fusion unit, thereby obtaining an optimized fusion feature map. The target detection module 240 is configured to perform target detection using a detection head based on the optimized fusion feature map, thereby obtaining a detection result.

[0101] The specific limitations of the infrared image target detection device can be referred to the limitations of the infrared image target detection method described above, which will not be repeated here. The various modules in the above infrared image target detection device can be realized by software, hardware, and combinations thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.

[0102] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 14 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an infrared image target detection method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0103] Those skilled in the art will understand that Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0104] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: Acquire infrared images of the target to be detected; The infrared image is input into a feature extraction network for feature extraction to obtain a multi-scale feature map. The feature extraction network uses an adaptive weight partitioning unit for feature extraction. The adaptive weight partitioning unit achieves real-time selection of convolution kernels and adaptive adjustment of feature extraction by integrating a dynamic convolution kernel weighting mechanism. The deep-scale feature map in the multi-scale feature map is processed by a polarization dynamic attention fusion unit to obtain the processed deep-scale feature map. The polarization dynamic attention fusion unit realizes feature enhancement and optimization of the deep-scale feature map through multi-scale convolution, attention mechanism and dynamic feature regulation. The processed deep-scale feature map and other multi-scale feature maps are input into the adaptive hierarchical feature path aggregation network. The input feature maps at each scale are fused from top to bottom and from bottom to top across scales. The result of the cross-scale fusion is then fused and optimized with the input feature maps at each scale using the adaptive hierarchical feature fusion unit to obtain the optimized fused feature map. The detection head is used to detect the target according to the optimized fusion feature map, and a detection result is obtained.

[0105] In one embodiment, a computer readable storage medium is provided, and a computer program is stored on the computer readable storage medium. The computer program is executed by a processor to implement the following steps: An infrared image to be subjected to target detection is acquired; The infrared image is input into a feature extraction network to extract features, and a multi-scale feature map is obtained. In the feature extraction network, an adaptive weight division unit is used to extract features. The adaptive weight division unit realizes real-time selection and adaptive adjustment of convolution kernels for feature extraction through a dynamic convolution kernel weighting mechanism. Deep layer scale feature maps in the multi-scale feature map are processed using a polarization dynamic attention fusion unit to obtain processed deep layer scale feature maps. The polarization dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep layer scale feature maps through multi-scale convolution, attention mechanism and dynamic feature regulation. The processed deep layer scale feature maps and other multi-scale feature maps are input into an adaptive hierarchical feature path aggregation network. Each scale feature map of each input is subjected to top-down and low-to-up cross-scale fusion. The cross-scale fused result and the input scale feature map are fused and optimized using an adaptive hierarchical feature fusion unit to obtain an optimized fusion feature map. The detection head is used to detect the target according to the optimized fusion feature map, and a detection result is obtained.

[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0107] The technical features of the above embodiments can be combined in any way. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0108] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent protection of the present application should be subject to the appended claims.

Claims

1. An infrared image target detection method, characterized by, The method comprises: acquiring an infrared image to be subjected to target detection; inputting the infrared image into a feature extraction network to extract features, to obtain a multi-scale feature map, wherein an adaptive weight division unit is used in the feature extraction network to extract features, the adaptive weight division unit realizes real-time selection and adaptive adjustment of a convolution kernel by fusing a dynamic convolution kernel weighting mechanism; processing deep layer scale feature maps in the multi-scale feature map by using a polarized dynamic attention fusion unit to obtain processed deep layer scale feature maps, the polarized dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep layer scale feature maps by multi-scale convolution, attention mechanism and dynamic feature regulation; inputting the processed deep layer scale feature maps and other multi-scale feature maps into an adaptive hierarchical feature path aggregation network, performing top-down and low-to-up cross-scale fusion on each scale feature map of each input, and performing fusion optimization on the cross-scale fused results and each scale feature map of the input by using an adaptive hierarchical feature fusion unit to obtain an optimized fused feature map; performing target detection according to the optimized fused feature map by using a detection head to obtain a detection result.

2. The infrared image target detection method of claim 1, wherein, In the adaptive weight division unit: the input image is respectively split into two parts and respectively input into two parallel adaptive weight division convolution branches, in the adaptive weight division convolution branch, the input data is differentially extracted by adaptive average pooling, dynamic kernel weight generation and weighted convolution, and then the feature expression is enhanced by splicing and fusion; the data output by the two parallel adaptive weight division convolution branches is spliced, and then a convolution layer is used to obtain a feature map of a certain scale.

3. The infrared image target detection method of claim 2, wherein, The feature extraction network comprises a convolution layer and a plurality of feature extraction blocks composed of a convolution layer and the adaptive weight division unit; the input image is sequentially input into a plurality of feature extraction blocks after passing through a convolution layer, and different scale feature maps are output by the adaptive weight division units in the feature extraction blocks.

4. The infrared image target detection method of claim 2, wherein, The polarized dynamic attention fusion unit takes the deep layer scale feature map and the corresponding position encoding as dual-channel input, first uses the Polo linear attention mechanism to perform long-range association, uses the ADD block to realize feature accumulation, uses the DyT block to dynamically regulate the features, uses the reinforcement channel attention block to focus on key channel features, and finally integrates by using the Mona block.

5. The infrared image target detection method of claim 4, wherein, In the Mona block: the input data is first normalized and then reduced in dimension by a Down Projection layer; the reduced data is divided into multiple paths, and different depth convolution layers are used to extract corresponding scale features, the features of multiple scales are fused by an Average layer, and residual fusion is performed with the reduced data; the obtained data after fusion is further subjected to convolution, GeLU activation, Up Projection, and dimension increasing, wherein, after the convolution, intermediate residual fusion is further performed with the reduced data; The data after dimensionality increase is fused with the input data in residual, and multi-scale feature deep integration and enhancement are completed.

6. The infrared image target detection method of claim 4, wherein, After the reinforced channel attention block is connected behind the Mona block, in the reinforced channel attention block: A double-branch processing mode is adopted to process the features obtained after the Mona block and the features not obtained after the Mona block, wherein the features obtained after the Mona block are processed through a branch for conventional feature transformation, and the features not obtained after the Mona block are processed through another branch for multi-scale convolution and up-sampling to mine features; After the output features of the two branches are spliced and fused, the intelligent weighting is performed by using the gating mechanism to realize the fine integration and enhancement of the features.

7. The infrared image target detection method of claim 1, wherein, In the adaptive hierarchical feature fusion unit: Two input features are respectively reduced to obtain first reduced features and second reduced features; After the first reduced features and the second reduced features are added, a convolution layer is used to realize preliminary fusion to generate baseline features; The first reduced features and the second reduced features are respectively processed by the local-global attention block to obtain first optimized features and second optimized features, wherein the local-global attention block includes two parallel local and global attention branches; The baseline features, the first optimized features and the second optimized features are spliced and fused, and then sequentially pass through a composite block composed of a convolution layer, a reparameterization convolution layer and a convolution layer (Conv1×1, RepConv3×3, Conv1×1) to obtain the output features of the adaptive hierarchical feature fusion unit.

8. An infrared image target detection apparatus characterized by comprising: The device comprises: An image acquisition module is configured to acquire an infrared image to be subjected to target detection. A multi-scale feature extraction module is configured to input the infrared image into a feature extraction network to extract features and obtain multi-scale feature maps, wherein the feature extraction network adopts an adaptive weight division unit to extract features, and the adaptive weight division unit realizes real-time selection and adaptive adjustment of convolution kernels by fusing a dynamic convolution kernel weighting mechanism. A deep-scale feature map optimization module is configured to process deep-scale feature maps in the multi-scale feature maps by using a polarization dynamic attention fusion unit to obtain processed deep-scale feature maps, and the polarization dynamic attention fusion unit realizes feature enhancement and optimization processing of the deep-scale feature maps by using multi-scale convolution, attention mechanism and dynamic feature regulation. A multi-scale feature fusion module is configured to input the processed deep-scale feature maps and other multi-scale feature maps into an adaptive hierarchical feature path aggregation network, perform top-down and low-to-up cross-scale fusion on each scale feature map, and fuse the cross-scale fused results with each scale feature map by using an adaptive hierarchical feature fusion unit to obtain optimized fused feature maps. A target detection module is configured to perform target detection according to the optimized fused feature maps by using a detection head to obtain detection results. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 7. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images