Infrared target detection method and system based on dynamic feature cooperation and state space

Abstract

This invention provides an infrared target detection method and system based on dynamic feature collaboration and state space, relating to the field of UAV infrared image processing technology. The method specifically includes: continuously acquiring raw infrared images containing target regions and preprocessing them to obtain a preprocessed infrared image sequence; constructing and training an infrared target detection model based on dynamic feature collaboration and state space to obtain a trained infrared target detection model based on dynamic feature collaboration and state space; inputting the preprocessed infrared image sequence into the trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection result. This invention combines dynamic feature collaboration with state space transformation to form a hierarchical feature refinement architecture, aiming to solve the problems of low signal-to-noise ratio, thermal crossover, and scale variation in infrared images.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing technology for unmanned aerial vehicles (UAVs), and in particular to an infrared target detection method and system based on dynamic feature collaboration and state space. Background Technology

[0002] Infrared imaging has become an indispensable sensing method due to its ability to capture the thermal radiation of a target. Unlike visible light imaging, infrared cameras remain effective in low-light, inclement weather, and partially obstructed conditions, thus reducing blind spots in safety-critical scenarios such as security monitoring, emergency rescue, and autonomous monitoring. Despite its practical value, target detection in infrared images remains challenging. Compared to visible light images, infrared images typically exhibit weak target-background contrast, sparse texture, and thermal clutter, where target responses are often intertwined with complex background radiation. Furthermore, targets captured in low-altitude UAV scenarios often undergo large-scale changes due to variations in imaging distance and altitude, and many targets exhibit weak features of only a few pixels. Optimization methods for visible light images cannot be directly applied to infrared images. The main reason is the inability to capture the smooth radiation patterns and low signal-to-noise ratio structure of infrared images. Summary of the Invention

[0003] To address the shortcomings of the existing technologies, this invention combines dynamic feature collaboration with state space transformation to form a hierarchical feature refinement architecture that integrates a dynamic calibration region attention module, an attention scale sequence fusion module, a visual state space module, and a content-guided triple attention fusion module. It proposes an infrared target detection method and system based on dynamic feature collaboration and state space, aiming to solve the problems of low signal-to-noise ratio, thermal crossover, and scale variation in infrared images, and achieve high-precision, low-latency, and robust detection of small targets on UAVs.

[0004] On the one hand, this invention proposes an infrared target detection method based on dynamic feature collaboration and state space, comprising:

[0005] Continuously acquire raw infrared images containing the target region and their corresponding real labels. Preprocess all acquired raw infrared images and associate them with their corresponding real labels to construct a training set.

[0006] An infrared target detection model based on dynamic feature collaboration and state space is constructed and trained using a training set to obtain a trained infrared target detection model based on dynamic feature collaboration and state space.

[0007] The original infrared image sequence to be detected is acquired and preprocessed. The preprocessed infrared image sequence is then input into a trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection result.

[0008] Furthermore, the specific method for preprocessing the continuously acquired raw infrared images containing the target region is as follows:

[0009] Perform the following operations frame-by-frame on the continuously acquired raw infrared images containing the target region:

[0010] Blind pixel replacement and non-uniformity correction are performed on the original infrared image of the current frame to smooth fixed pattern noise and obtain the corrected infrared image.

[0011] Adaptive histogram equalization or automatic gain control algorithm is used to perform dynamic range compression and linear stretching on the corrected infrared image to obtain the compressed infrared image.

[0012] The compressed infrared image is subjected to channel expansion and size normalization to obtain a preprocessed infrared image, which in turn yields a preprocessed infrared image sequence.

[0013] Furthermore, the infrared target detection model based on dynamic feature collaboration and state space includes: a backbone network, a neck network, and a detection head;

[0014] The backbone network is used to extract features from the preprocessed infrared image and output a multi-scale feature map.

[0015] The neck network is used to perform dynamic feature collaboration and state space transformation on multi-scale feature maps and output multi-scale fused feature maps.

[0016] The detection head is used to perform regression and classification prediction on the multi-scale fused feature map in parallel, and to perform non-maximum suppression processing on the obtained prediction box, target category and confidence score to obtain the final infrared target detection result.

[0017] The infrared target detection results include: detection bounding box, target category, and confidence level.

[0018] Furthermore, the backbone network includes, in sequence: a convolutional block, a first convolutional layer, a first cross-stage parallel convolutional module, a second convolutional layer, a second cross-stage parallel convolutional module, a third convolutional layer, a first dynamic calibration region attention module, a fourth convolutional layer, and a second dynamic calibration region attention module.

[0019] The convolutional block is used to sequentially perform convolution, batch normalization, and SiLU activation function operations on the input preprocessed infrared image to obtain an initial feature map;

[0020] The first, second, third, and fourth convolutional layers are all independent convolutional layers;

[0021] The first and second cross-stage parallel convolutional modules have the same network architecture, both including: a standard convolutional layer, a channel splitting layer, a backbone computation branch, a cross-stage direct connection branch, a channel concatenation layer, and an output layer. The standard convolutional layer maps the input feature map to a preset channel dimension. The channel splitting layer equally divides the mapped feature map into a first branch feature map and a second branch feature map along the channel dimension. The backbone computation branch sequentially performs standard convolution, custom receptive field convolution, and deconvolution on the first branch feature map to obtain a deep feature map. The cross-stage direct connection branch directly passes the second branch feature map to the channel concatenation layer. The channel concatenation layer concatenates the deep feature map and the second branch feature map along the channel dimension. The output layer performs channel fusion and dimensionality reduction on the concatenated feature map to output a cross-stage fused feature map.

[0022] The output of the first cross-stage parallel convolution module is the first cross-stage fused feature map; the output of the second cross-stage parallel convolution module is the second cross-stage fused feature map.

[0023] The first and second dynamic calibration region attention modules have the same network architecture, both including: a dynamic calibration unit for performing lightweight dynamic calibration on the input feature map through dynamic hyperbolic tangent operation; and a dense aggregation unit for performing multi-level feature extraction and fusion on the calibrated feature map to obtain a feature representation. Residual connection units are used for feature representation. After channel-by-channel scaling, the refined feature map is added to the input feature map to obtain the refined feature map.

[0024] The output of the first dynamic calibration region attention module is a first refined feature map; the output of the second dynamic calibration region attention module is a second refined feature map.

[0025] The first cross-stage fusion feature map, the second cross-stage fusion feature map, the first refined feature map, and the second refined feature map are used as multi-scale feature maps.

[0026] Further, the neck network includes:

[0027] The attention scale sequence fusion module is used to perform dimensional matching and fusion of multi-scale feature maps to obtain a scale fusion feature map; it performs two-layer progressive feature enhancement and state space modeling on the multi-scale feature maps to obtain a first-level state feature map and a second-level state feature map; and then uses a one-dimensional convolutional channel attention mechanism to align and weight the scale fusion feature map and the second-level state feature map to obtain a first fusion feature map.

[0028] The first splicing layer is used to splice the first-level state feature map and the second-level state feature map to obtain the first spliced ​​feature map;

[0029] The first visual state space module is used to perform visual state space modeling on the first stitched feature map to obtain the global state feature map.

[0030] The first triple attention fusion module is used to adaptively weight the first fused feature map with the global state feature map in the channel dimension, spatial dimension and pixel dimension to obtain the second fused feature map.

[0031] The second concatenation layer is used to concatenate the second refined feature map in the multi-scale feature map with the global state feature map to obtain the second concatenated feature map.

[0032] The third cross-stage parallel convolution module is used to perform channel mapping on the second concatenated feature map and equally divide it into two branch feature maps. The first branch feature map is then restored by standard convolution, custom receptive field convolution, and deconvolution. After being concatenated with the second branch feature map, it is fused by channels and reduced in dimension to obtain the third cross-stage fused feature map.

[0033] The second and third attention fusion module is used to adaptively weight the second refined feature map and the third cross-stage fusion feature map in the multi-scale feature map in the channel dimension, spatial dimension and pixel dimension to obtain the third fusion feature map.

[0034] The first fusion feature map, the second fusion feature map, and the third fusion feature map are used as multi-scale fusion feature maps.

[0035] Furthermore, the attention scale sequence fusion module includes: a scale sequence fusion module, a first triple feature enhancement module, a second visual state space module, a second triple feature enhancement module, a third visual state space module, and an attention guidance calibration module;

[0036] The scale sequence fusion module is used to project the second cross-stage fused feature map, the first refined feature map, and the second refined feature map in the multi-scale feature map to the same resolution and stack them along the channel dimension. The stacked feature map is then subjected to three-dimensional batch normalization, nonlinear activation, and three-dimensional max pooling in sequence to obtain the scale fused feature map.

[0037] The first triple feature enhancement module is used to enhance the second cross-stage fusion feature map, the first refined feature map and the second refined feature map in the multi-scale feature map respectively, and to stitch the enhanced feature maps along the channel dimension to obtain the first-level enhanced feature map.

[0038] The second visual state space module is used to perform visual state space modeling on the first-level enhanced feature map to obtain the first-level state feature map.

[0039] The second and third feature enhancement modules are used to enhance the first cross-stage fusion feature map, the second cross-stage fusion feature map, and the first-level state feature map in the multi-scale feature map respectively, and to stitch the enhanced feature maps along the channel dimension to obtain the second-level enhanced feature map.

[0040] The third visual state space module is used to perform visual state space modeling on the secondary enhanced feature map to obtain the secondary state feature map.

[0041] The attention-guided calibration module performs channel calibration on the secondary state feature map using a one-dimensional convolutional channel attention mechanism, and then sums it pixel-by-pixel with the scale fusion feature map to obtain residual fusion features; then, it uses a position attention component to perform spatial attention weighting on the residual fusion features to obtain the first fusion feature map.

[0042] Furthermore, the first triple feature enhancement module and the second triple feature enhancement module have the same network architecture, both including: defining the input three feature maps as high-resolution feature maps, medium-resolution feature maps and low-resolution feature maps in order of resolution from high to low;

[0043] Adaptive max pooling and average pooling are performed simultaneously on the high-resolution feature map, and the resulting max pooled features and average pooled features are added element by element to obtain the high-resolution enhanced feature map.

[0044] Nearest neighbor interpolation is performed on the medium-resolution feature map and the low-resolution feature map respectively to obtain the medium-resolution enhanced feature map and the low-resolution enhanced feature map;

[0045] The obtained high-resolution enhanced feature map, medium-resolution enhanced feature map and low-resolution enhanced feature map are concatenated along the channel dimension to obtain the output enhanced feature map;

[0046] The enhanced feature map output by the first triple feature enhancement module is a first-level enhanced feature map; the enhanced feature map output by the second triple feature enhancement module is a second-level enhanced feature map.

[0047] Furthermore, the first visual state space module, the second visual state space module, and the third visual state space module have the same network architecture, all including:

[0048] The layer normalization unit is used to perform layer normalization on the input feature map to obtain the normalized feature map;

[0049] The first depthwise convolutional unit is used to perform depthwise separable convolution on the normalized feature map to obtain a locally enhanced feature map.

[0050] The hidden state mixer's state space layer is used to capture long-range dependencies in the local augmented feature map using a discrete state space model, and outputs a state augmented feature map.

[0051] The feature projection layer is used to perform feature projection and batch normalization on the state-enhanced feature map, and to project and normalize the feature map.

[0052] The second depthwise convolutional unit is used to perform depthwise separable convolution on the projection normalized feature map to obtain a depth-enhanced feature map.

[0053] The feedforward network unit is used to perform inter-channel information exchange on the deep augmented feature map and output the channel interaction feature map.

[0054] The hierarchical scaling unit is used to multiply the channel interaction feature map with a learnable diagonal weight matrix to obtain a scaled feature map.

[0055] The residual connection unit is used to add the scaled feature map to the input feature map element by element to obtain the state feature map;

[0056] The state feature map output by the first visual state space module is a global state feature map; the state feature map output by the second visual state space module is a first-level state feature map; and the state feature map output by the third visual state space module is a second-level state feature map.

[0057] Furthermore, the first and second triple attention fusion modules have the same network architecture, both including: defining the lower-resolution feature map in the two input feature maps as a high-level feature map. High-resolution feature maps are defined as low-level features. ;

[0058] Initial aggregation unit, used to aggregate high-level features and low-level features Initial fusion is performed through element-wise addition to obtain the initial aggregated features. ;

[0059] Channel attention branch, used for initial aggregated features Calculate channel attention to obtain channel attention features;

[0060] Spatial attention branch, used for initial aggregated features Calculate spatial attention to obtain spatial attention features;

[0061] The splicing and reshaping unit is used to combine channel attention features, spatial attention features, and initial aggregated features. The concatenated features are then reshaped using a tensor reshaping operation. The reshaped features are then grouped and convolved, and finally, content-adaptive spatial weights are generated using a sigmoid activation function. ;

[0062] Weighted fusion unit, used to utilize content-adaptive spatial weights For high-level features and low-level features Perform weighted fusion to combine the initial aggregated features The features are connected to the weighted fused feature residuals, and the channel dimension of the residual-connected features is optimized by pointwise convolution to output a fused feature map.

[0063] The fusion feature map output by the first triple attention fusion module is used as the second fusion feature map; the fusion feature map output by the second triple attention fusion module is used as the third fusion feature map.

[0064] On the other hand, this invention proposes an infrared target detection system based on dynamic feature cooperation and state space, the system comprising:

[0065] The training set construction module is used to continuously acquire raw infrared images containing the target region and their corresponding real labels, preprocess all acquired raw infrared images and associate them with their corresponding real labels to construct a training set.

[0066] The detection model building module is used to build and train an infrared target detection model based on dynamic feature collaboration and state space to obtain a trained infrared target detection model based on dynamic feature collaboration and state space.

[0067] The infrared image acquisition module is used to acquire the original infrared image sequence to be detected and to preprocess it to obtain the preprocessed infrared image sequence.

[0068] The infrared target detection module is used to input the preprocessed infrared image sequence into the trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection results.

[0069] The beneficial effects of adopting the above technical solution are as follows:

[0070] This invention combines dynamic feature collaboration with state space transformation to propose a target detection model based on a hierarchical feature refinement architecture. Specifically, a dynamic range calibration region attention module is introduced into the backbone network to stabilize feature activation under strong thermal clutter. In the neck network, an attention scale sequence fusion strategy reduces cross-scale semantic misalignment and ensures accurate spatial coherence. Furthermore, a visual state space module captures global contextual dependencies while maintaining linear computational complexity. Finally, a content-guided triple attention fusion module maximizes feature discriminability by calibrating the fusion representation across channel, spatial, and pixel dimensions. This model exhibits enhanced robustness under degraded imaging conditions, making it highly suitable for practical applications such as aerial surveillance and autonomous monitoring.

[0071] To address the issues of low signal-to-noise ratio, thermal crossover, and scale variations in infrared images, the proposed target detection model integrates four specific modules: a dynamic calibration region attention module for adaptive feature enhancement to stabilize feature activation under strong thermal clutter; an attention-scale sequence fusion module for achieving accurate cross-layer alignment and scale-consistent multi-level feature fusion, reducing cross-scale semantic misalignment; a visual state space module for optimizing linearly complex state representations and capturing long-range contextual dependencies, thereby capturing global contextual dependencies while maintaining linear computational complexity; and a content-guided triple attention fusion module that maximizes feature discriminability by calibrating fusion representations across channel, spatial, and pixel dimensions, achieving fine-grained feature optimization across multiple dimensions. This model can reduce the false negative rate of weak thermal signals in complex and densely populated backgrounds while maintaining the real-time inference efficiency required by UAVs.

[0072] To address the semantic misalignment and thermal clutter issues in weak infrared target detection, this invention combines linear-complexity global inference with an alignment-first scale fusion strategy, proposing a method that integrates dynamic feature calibration and scale-consistent fusion. By explicitly coupling a dynamic calibration region attention module and an attention-scale sequence fusion module, the problem of weak target activation and scale misalignment in complex infrared scenes is effectively solved.

[0073] This invention introduces a global reasoning and multidimensional calibration paradigm. By integrating a visual state space module and a content-guided triple attention fusion module, it successfully models long-range dependencies and improves the discriminability of targets. Simultaneously, this invention avoids the high computational cost of standard self-attention. Attached Figure Description

[0074] Figure 1 This is a flowchart of the infrared target detection method based on dynamic feature collaboration and state space in this embodiment;

[0075] Figure 2This is a network architecture diagram of the infrared target detection model based on dynamic feature collaboration and state space in this embodiment;

[0076] Figure 3 This is a schematic diagram of the attention scale sequence fusion module in this embodiment.

[0077] Figure 4 This is a schematic diagram of the triple attention fusion module in this embodiment;

[0078] Figure 5 This is a structural diagram of the infrared target detection system based on dynamic feature collaboration and state space in this embodiment. Detailed Implementation

[0079] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0080] Example 1:

[0081] The infrared target detection method based on dynamic feature collaboration and state space in this embodiment, such as Figure 1 As shown, the method includes the following steps:

[0082] The raw infrared images containing the target region and their corresponding ground truth labels are continuously acquired. All acquired raw infrared images are preprocessed and associated with their corresponding ground truth labels to construct a training set.

[0083] In this embodiment, by integrating infrared photoelectric sensing devices onto the three-axis stabilized gimbal of the UAV, it is possible to perform top-down or high-angle oblique-view detection of ground or low-altitude targets from the air, thereby acquiring continuous infrared images containing the target area. Compared with visible light images, the acquired infrared images have the following significant physical characteristics: First, they lack color and texture, and the pixel grayscale values ​​only reflect the differences in scene temperature and emissivity; second, they have low contrast and low signal-to-noise ratio, and are generally dark and have blurred edges due to atmospheric attenuation and detector thermal noise; third, they suffer from severe background clutter interference, and under the UAV's top-down view, hot objects on the ground can easily form false alarm sources similar to real targets; fourth, they contain blind pixels and non-uniform stripe noise.

[0084] Therefore, raw infrared data cannot be directly input into mainstream deep learning target recognition networks for processing; it must undergo a rigorous preprocessing stage. This is because the raw data has an excessively wide dynamic range, while the actual temperature difference of a target often occupies only a very narrow grayscale range. Without dynamic range compression and contrast enhancement, the target's subtle features will be overwhelmed by high-frequency background information. Furthermore, the inherent blind cells of the detector exhibit extremely high local gradients during convolution operations, making them highly susceptible to misidentification by the network as "small targets," leading to severe false alarms.

[0085] The real labels include the target bounding box and the category label.

[0086] The specific method for preprocessing continuously acquired raw infrared images containing the target region is as follows:

[0087] Perform the following operations frame-by-frame on the continuously acquired raw infrared images containing the target region:

[0088] Blind pixel replacement and non-uniformity correction are performed on the original infrared image of the current frame to smooth the fixed pattern noise and obtain the corrected infrared image.

[0089] In this embodiment, blind pixel detection is performed on any frame of the original infrared image, and the pixels are replaced based on the pixel values ​​of the effective neighborhood surrounding the blind pixel. For example, the nearest neighbor substitution method can be used to replace the blind pixel with the gray value of the adjacent effective pixel, or the neighborhood averaging method can be used to compensate for the blind pixel by using the average value of the eight effective pixels surrounding the blind pixel. Then, the replaced infrared image is subjected to non-uniformity correction, such as using a two-point correction method or a time-domain high-pass filtering method, to eliminate fixed pattern noise.

[0090] Adaptive histogram equalization or automatic gain control algorithm is used to dynamically compress and linearly stretch the corrected infrared image to highlight the target contour, thus obtaining a compressed infrared image.

[0091] The compressed infrared image is channel expanded and its size normalized to convert it into tensor data that conforms to the input standard of the target detection model, thus obtaining a preprocessed infrared image, and then a preprocessed infrared image sequence.

[0092] In this embodiment, the size of the original infrared image is 640×640 pixels, and the size of the infrared image output after preprocessing remains unchanged at 640×640 pixels.

[0093] An infrared target detection model based on dynamic feature collaboration and state space is constructed and trained using a training set to obtain a trained infrared target detection model based on dynamic feature collaboration and state space.

[0094] like Figure 2As shown, the infrared target detection model based on dynamic feature collaboration and state space includes: a backbone network, a neck network, and a detection head.

[0095] The backbone network is used to extract features from the preprocessed infrared image and output a multi-scale feature map.

[0096] The backbone network comprises, in sequence: a convolutional block, a first convolutional layer, a first cross-stage parallel convolutional module, a second convolutional layer, a second cross-stage parallel convolutional module, a third convolutional layer, a first dynamic calibration region attention module, a fourth convolutional layer, and a second dynamic calibration region attention module.

[0097] The convolutional block is used to sequentially perform convolution, batch normalization, and SiLU activation function operations on the input preprocessed infrared image to obtain an initial feature map;

[0098] The first, second, third, and fourth convolutional layers are independent convolutional layers used to perform convolution operations on the input feature map to adjust the number of channels or spatial size.

[0099] The first and second cross-stage parallel convolutional modules have the same network architecture, both including: a standard convolutional layer, a channel splitting layer, a backbone computation branch, a cross-stage direct connection branch, a channel concatenation layer, and an output layer. The standard convolutional layer maps the input feature map to a preset channel dimension. The channel splitting layer equally divides the mapped feature map into a first branch feature map and a second branch feature map along the channel dimension. The backbone computation branch sequentially performs standard convolution, custom receptive field convolution, and deconvolution on the first branch feature map to obtain a deep feature map. The cross-stage direct connection branch directly passes the second branch feature map to the channel concatenation layer. The channel concatenation layer concatenates the deep feature map and the second branch feature map along the channel dimension. The output layer performs channel fusion and dimensionality reduction on the concatenated feature map to output a cross-stage fused feature map.

[0100] The output of the first cross-stage parallel convolution module is the first cross-stage fused feature map; the output of the second cross-stage parallel convolution module is the second cross-stage fused feature map.

[0101] In this embodiment, the cross-stage parallel convolution module, by introducing a bottleneck layer structure with a variable receptive field, can take into account both local micro-edges and macroscopic background context information when extracting infrared target features. This module uses convolution and deconvolution to extract features at different stages, and concatenates the branches to obtain a feature map. Specifically, it employs... The standard convolutional layer integrates the channel information of the input features and maps the number of channels to the channel dimension defined by the network. The mapped feature map is then equally divided into a first branch feature map and a second branch feature map along the channel dimension, and each branch passes through a backbone computation branch and a cross-stage direct connection branch, respectively. The second branch feature map, after passing through the cross-stage direct connection branch, undergoes no additional convolutional operations and is directly mapped to the end of the module to preserve the original shallow / fine-grained feature information and prevent network degradation. The backbone computation branch is used for deep feature extraction, including: Standard convolution (for channel reduction), one A specific sequence consisting of a custom receptive field convolution (for extracting local small edge features) and a deconvolution layer for feature reconstruction is used for feature extraction at different stages. The custom receptive field convolution includes: a stride of 2 and a convolution kernel of... The sliding convolution, batch normalization, and SiLU activation function are used. The SiLU activation function can be expressed as:

[0102]

[0103] in, This indicates that the input feature values ​​come from the feature map after batch normalization. This represents the output of the SiLU activation function; For the Sigmoid function; It is a natural constant.

[0104] After the main computation branch completes feature extraction, its output features are concatenated with the feature map of the second branch connected across stages in the channel dimension. The number of channels in the concatenated feature map will increase significantly. Then, it is processed through a... The convolutional layers perform cross-channel information fusion. It should be noted that the kernel size and stride in the cross-stage parallel convolution module can be adaptively adjusted according to the actual application scenario.

[0105] The first and second dynamic calibration region attention modules have the same network architecture, both including: a dynamic calibration unit for performing lightweight dynamic calibration on the input feature map through dynamic hyperbolic tangent operation; and a dense aggregation unit for performing multi-level feature extraction and fusion on the calibrated feature map to obtain a feature representation. Residual connection units are used for feature representation. After channel-by-channel scaling, the refined feature map is added to the input feature map to obtain the refined feature map.

[0106] In this embodiment, the feature map of the input dynamic calibration region attention module... , Indicates the number of channels. Indicates spatial height, To represent the spatial width, a lightweight dynamic calibration is first performed using the Dynamic Hyperbolic Tangent (DyT) operation, expressed as:

[0107]

[0108] in, The calibrated feature map; These are channel-level scaling parameters that can be dynamically learned. ; This indicates element-wise multiplication; Let represent the hyperbolic tangent function, and It is a nonlinear function used to effectively suppress anomalous activation while preserving informative feature changes within its approximately linear central region; For learnable global scaling parameters, ; These are channel-level offset parameters that can be dynamically learned. To achieve stable optimization, it is usually set that... and The initial value is 1. The initial value is 0.

[0109] The calibrated feature map Through a The convolutional layers are projected into a dense aggregate structure, which specifically involves processing the calibrated feature maps through a series of region attention modules. The process involves sequential processing and concatenating all the intermediate features to preserve low-level thermal noise details and high-level semantic context. This process can be represented as follows:

[0110]

[0111]

[0112]

[0113] in, This represents the projected feature map; Indicates the use of Projection operations are performed on the convolutional layers; This represents the number of region attention modules, set to [number] in this embodiment. ; Indicates the first The output of each region's attention module; Indicates the first The input to the attention module of each region; This indicates the processing operations of the region attention module; This indicates a splicing operation.

[0114] The residual connection units, enhanced by layer scaling, ensure training stability, thereby obtaining refined feature maps. , is represented as:

[0115]

[0116] in, This is a learnable scaling factor for each channel. This embodiment will The initial value is set to 1.

[0117] In this embodiment, the dynamic range calibration region attention module is used to dynamically calibrate and stably extract weak thermal features. The parameterized Tanh function (i.e., dynamic hyperbolic tangent operation) enhances the sensitivity to subtle thermal changes by dynamically adjusting the feature representation through optimized learnable variables.

[0118] The output of the first dynamic calibration region attention module is a first refined feature map; the output of the second dynamic calibration region attention module is a second refined feature map.

[0119] The first cross-stage fusion feature map, the second cross-stage fusion feature map, the first refined feature map, and the second refined feature map are used as multi-scale feature maps.

[0120] The neck network is used to perform dynamic feature collaboration and state space transformation on multi-scale feature maps, and output multi-scale fused feature maps.

[0121] The neck network includes:

[0122] The attention scale sequence fusion module is used to perform dimensional matching and fusion of multi-scale feature maps to obtain a scale fusion feature map; it performs two-layer progressive feature enhancement and state space modeling on the multi-scale feature maps to obtain a first-level state feature map and a second-level state feature map; then, based on a one-dimensional convolutional channel attention mechanism, it aligns and weights the scale fusion feature map and the second-level state feature map to obtain a first fusion feature map.

[0123] In this embodiment, the attention scale sequence fusion module utilizes the attention scale sequence fusion strategy to achieve accurate cross-layer alignment and scale-consistent multi-level feature fusion.

[0124] The first splicing layer is used to splice the first-level state feature map and the second-level state feature map to obtain the first spliced ​​feature map.

[0125] It should be noted that, as Figure 2As shown, to ensure that the number of channels in the two feature maps input to the first stitching layer is consistent, this embodiment sets up a convolutional layer to align the two feature maps in the channel dimension before inputting them into the first stitching layer.

[0126] The first visual state space module is used to perform visual state space modeling on the first stitched feature map to obtain the global state feature map.

[0127] The first triple attention fusion module is used to adaptively weight the first fused feature map with the global state feature map in the channel dimension, spatial dimension and pixel dimension to obtain the second fused feature map.

[0128] The second concatenation layer is used to concatenate the second refined feature map in the multi-scale feature map with the global state feature map to obtain the second concatenated feature map.

[0129] It should be noted that, as Figure 2 As shown, to ensure that the number of channels in the two feature maps input to the second stitching layer is consistent, this embodiment sets up a convolutional layer to align the two feature maps in the channel dimension before inputting them into the second stitching layer.

[0130] The third cross-stage parallel convolution module is used to perform channel mapping on the second concatenated feature map and equally divide it into two branch feature maps. The first branch feature map is then restored by standard convolution, custom receptive field convolution, and deconvolution. After being concatenated with the second branch feature map, it undergoes channel fusion and dimensionality reduction to obtain the third cross-stage fused feature map.

[0131] In this embodiment, the third cross-stage parallel convolution module has the same network architecture as the first cross-stage parallel convolution module and the second cross-stage parallel convolution module.

[0132] The second and third attention fusion modules are used to adaptively weight the second refined feature map and the third cross-stage fusion feature map in the multi-scale feature map in the channel dimension, spatial dimension and pixel dimension to obtain the third fusion feature map.

[0133] The first fusion feature map, the second fusion feature map, and the third fusion feature map are used as multi-scale fusion feature maps.

[0134] like Figure 3 As shown, the attention scale sequence fusion module includes: a scale sequence fusion module, a first triple feature enhancement module, a second visual state space module, a second triple feature enhancement module, a third visual state space module, and an attention guidance calibration module.

[0135] The scale sequence fusion module is used to project the second cross-stage fused feature map, the first refined feature map, and the second refined feature map in the multi-scale feature map to the same resolution and stack them along the channel dimension. The stacked feature map is then subjected to three-dimensional batch normalization, nonlinear activation, and three-dimensional max pooling in sequence to obtain the scale fused feature map.

[0136] In this embodiment, the scale sequence fusion module first addresses the feature mismatch problem by projecting heterogeneous feature maps—the second cross-stage fused feature map, the first refined feature map, and the second refined feature map—to a shared reference resolution. Then, each projected feature map undergoes dimensional expansion and is stacked and aggregated along the newly introduced scale axis (i.e., channel dimension). Finally, by combining 3D batch normalization, the LeakyReLU activation function (i.e., non-linear activation), and 3D max pooling, salient feature representations across scales, i.e., the scale fused feature map, are extracted. .

[0137]

[0138] in, This represents a three-dimensional max pooling operation, which is adopted in this embodiment. ; This represents the LeakyReLU activation function; This indicates a three-dimensional batch normalization operation; This embodiment uses the three-dimensional convolution factor. ; This represents a tensor stacking operation along the channel dimension; Represents a series of input two-dimensional feature maps. In this embodiment, the number of input feature maps is [number]. .

[0139] The first triple feature enhancement module is used to enhance the second cross-stage fusion feature map, the first refined feature map, and the second refined feature map in the multi-scale feature map respectively, and to stitch the enhanced feature maps along the channel dimension to obtain the first-level enhanced feature map.

[0140] The second visual state space module is used to perform visual state space modeling on the first-level enhanced feature map to obtain the first-level state feature map.

[0141] The second and third feature enhancement modules are used to enhance the first cross-stage fusion feature map, the second cross-stage fusion feature map, and the first-level state feature map in the multi-scale feature map, respectively, and then stitch the enhanced feature maps along the channel dimension to obtain the second-level enhanced feature map.

[0142] The third visual state space module is used to perform visual state space modeling on the secondary enhanced feature map to obtain the secondary state feature map.

[0143] The attention-guided calibration module performs channel calibration on the secondary state feature map using a one-dimensional convolutional channel attention mechanism, and then sums it pixel-by-pixel with the scale fusion feature map to obtain residual fusion features; then, it uses a position attention component to perform spatial attention weighting on the residual fusion features to obtain the first fusion feature map.

[0144] In this embodiment, the attention-guided calibration module employs a sequence residual design, which can effectively fuse the scale-fused feature map output by the scale sequence fusion module. The module uses a second-level state feature map output by the third-vision state space module to explicitly model cross-channel and spatial dependencies. First, a one-dimensional convolutional channel attention mechanism is applied to the second-level state feature map. Then, the resulting channel-calibrated features are combined with the scale-fused feature map via residual connections. Residual fusion is performed, and the fused representation captures orthogonal spatial prior information through a location attention component. The process is represented as follows:

[0145]

[0146] in, This represents the features after deep fusion by the attention-guided calibration module, i.e., the first fused feature map; Represents the second-level state feature map; This indicates a pixel-by-pixel summation operation; This represents the processing procedure of the one-dimensional convolutional channel attention mechanism; This represents the location attention component, which is used to capture orthogonal spatial prior information, that is, to model the orthogonal dependencies between different spatial locations in the feature map.

[0147] The first triple feature enhancement module and the second triple feature enhancement module have the same network architecture, both of which include: defining the input three feature maps as high-resolution feature maps, medium-resolution feature maps and low-resolution feature maps in order of resolution from high to low.

[0148] In this embodiment, for any triple feature enhancement module, the three feature maps input to the module are arranged from shallow to deep, with their downsampling ratios increasing sequentially. The shallow output feature with the smallest downsampling ratio (e.g., 8 times the downsampling ratio relative to the input image), the largest spatial size, and rich low-level spatial details is defined as a high-resolution feature map. The deep output feature with the largest downsampling ratio (e.g., 32 times the downsampling ratio relative to the input image), the smallest spatial size, and high-level semantic context is defined as a low-resolution feature map. Feature maps with downsampling ratios and spatial sizes between the two are defined as medium-resolution feature maps.

[0149] Adaptive max pooling and average pooling are performed simultaneously on the high-resolution feature map, and the resulting max pooled features and average pooled features are added element-wise to obtain a high-resolution enhanced feature map.

[0150] Nearest neighbor interpolation is performed on the medium-resolution feature map and the low-resolution feature map respectively to obtain the medium-resolution enhanced feature map and the low-resolution enhanced feature map.

[0151] The obtained high-resolution enhanced feature map, medium-resolution enhanced feature map, and low-resolution enhanced feature map are concatenated along the channel dimension to obtain the output enhanced feature map.

[0152] The enhanced feature map output by the first triple feature enhancement module is a first-level enhanced feature map; the enhanced feature map output by the second triple feature enhancement module is a second-level enhanced feature map.

[0153] In this embodiment, the triple feature enhancement module fuses local contextual information with fine details through resolution-aware resampling. For high-resolution feature maps, the module simultaneously performs adaptive max pooling and average pooling, followed by element-wise addition of their respective responses; for medium-resolution and low-resolution feature maps, nearest-neighbor interpolation is performed to accurately restore the spatial layout and reduce aliasing artifacts. Finally, all rescaled feature maps are stitched together along the channel dimension to generate a unified and scale-coordinated representation, i.e., the output enhanced feature map. .

[0154] It should be noted that, as Figure 2 As shown, to ensure that the number of channels of the three input feature maps of any of the three-way feature enhancement modules in the first and second triple feature enhancement modules is consistent, this embodiment sets up a convolutional layer to align the three input feature maps in the channel dimension before inputting them into the triple feature enhancement module.

[0155] The first visual state space module, the second visual state space module, and the third visual state space module have the same network architecture, all including:

[0156] The layer normalization unit is used to perform layer normalization on the input feature map to obtain the normalized feature map.

[0157] The first depthwise convolutional unit is used to perform depthwise separable convolution on the normalized feature map to obtain a locally enhanced feature map.

[0158] The hidden state mixer's state space layer is used to capture long-range dependencies in local augmented feature maps using a discrete state space model, and outputs state augmented feature maps.

[0159] The feature projection layer is used to perform feature projection and batch normalization on the state-enhanced feature map, and to project and normalize the feature map.

[0160] The second depthwise convolutional unit is used to perform depthwise separable convolution on the projection-normalized feature map to obtain a depth-enhanced feature map.

[0161] The feedforward network unit is used to perform inter-channel information exchange on the deep augmented feature map and output the channel interaction feature map.

[0162] The hierarchical scaling unit is used to multiply the channel interaction feature map with a learnable diagonal weight matrix to obtain a scaled feature map.

[0163] The residual connection unit is used to add the scaled feature map to the input feature map element by element to obtain the state feature map.

[0164] The state feature map output by the first visual state space module is a global state feature map; the state feature map output by the second visual state space module is a first-level state feature map; and the state feature map output by the third visual state space module is a second-level state feature map.

[0165] In this embodiment, the visual state space module, with linear computational complexity, is used to capture long-range contextual dependencies. This module utilizes the state-space dual layer of a Hidden State Mixer-based State SpaceDuality (HSM-SSD) to aggregate global information and is combined with a feedforward network specifically designed to handle inter-channel interactions. The state-space dual layer of the Hidden State Mixer employs a discrete state-space model to model these long-range dependencies, expressed as:

[0166]

[0167]

[0168] in, Indicates time The hidden state at that location; Indicates time The hidden state at that location; and These represent the discrete state-space model in time. The input and output sequences at the specified points; This is the state transition matrix; For the input projection matrix; To output the projection matrix; This is a direct matrix.

[0169] Specifically, the visual state space module ensures training stability through residual connections and hierarchical scaling. In the specific module structure topology, the data flow follows a design paradigm of pre-normalization and multi-level fusion: First, the input feature maps undergo layer normalization before entering the state space layer of the hidden state mixer to maintain numerical stability and prevent gradient anomalies in long sequence processing. Second, through... Depthwise separable convolutions are used to extract local features; in this embodiment, two convolutions are preferably configured. Depth-separable convolutional layers are symmetrically distributed in front of and behind the state-space dual layer of the hidden state mixer, namely the first depthwise convolutional unit and the second depthwise convolutional unit.

[0170] Batch normalization is applied in the feature projection or point convolution stage after the state space layer to optimize computational efficiency and accelerate model convergence. Subsequently, the features after state space hidden state modeling are transformed through a channel-wise feedforward network to achieve inter-channel communication. The computational flow of the feedforward network can be described as follows:

[0171]

[0172] in, This represents the final feature tensor output by the feedforward network after channel dimensionality upscaling, nonlinear activation, and dimensionality reduction and recombination, i.e., the channel interaction feature map; It is a non-linear activation function; This represents the intermediate feature tensor, i.e., the depth-enhanced feature map, input to the feedforward network module after the previous hidden state modeling process.

[0173] Next, hierarchical scaling is applied to the channel interaction feature map, that is, multiplying it by a learnable diagonal weight matrix to dynamically adjust the feature response intensity of the layer; finally, the scaled feature map is residually connected with the original input (i.e., element-wise addition) to complete the feature transfer and update of the entire module.

[0174] In this embodiment, an attention-scale sequence fusion strategy is employed to achieve precise cross-layer alignment and scale-consistent multi-level feature fusion. This strategy integrates a scale sequence module, a triple feature enhancement module, an attention-guided calibration module, and a visual state space module, forming a hierarchical feature refinement architecture. As a complement to global scale perception, the triple feature enhancement module aligns local contextual information and fine-grained details through resolution-aware resampling, thereby generating a unified representation robust to thermal noise.

[0175] like Figure 4 As shown, the first and second triple attention fusion modules have the same network architecture, both including: defining the lower-resolution feature map in the two input feature maps as a high-level feature map. High-resolution feature maps are defined as low-level features. .

[0176] In this embodiment, the content-guided triple attention fusion module collaboratively performs adaptive feature weighting across channel, spatial, and pixel dimensions. During cross-layer fusion, this module incorporates high-level features... and low-level features It is processed as input. Specifically, the high-level features Extracted from deeper layers in the object detection model, which have lower spatial resolution. Contains rich global semantic contextual information about the target and complex background; the low-level features It is extracted from a shallower layer in the target detection model, which has high spatial resolution and fully preserves the high-frequency edge morphology and fine spatial position details of small infrared targets.

[0177] Initial aggregation unit, used to aggregate high-level features and low-level features Initial fusion is performed through element-wise addition to obtain the initial aggregated features. .

[0178] In this embodiment, high-level features are added element-wise from adjacent levels. and low-level features The initial fusion is represented as:

[0179]

[0180] in, , This indicates a given step size. Considering the small size and susceptibility to background clutter of infrared targets, this embodiment uses a step size... The preferred setting is 8 or 16, which means that the feature space resolution of the fused image is 1 / 8 or 1 / 16 of the resolution of the original input image.

[0181] Channel attention branch, used for initial aggregated features Calculate channel attention to obtain channel attention features.

[0182] Spatial attention branch, used for initial aggregated features Calculate spatial attention to obtain spatial attention features.

[0183] In this embodiment, the channel attention branch captures global cross-channel dependencies through a squeeze-excitation mechanism. The spatial attention branch highlights salient target regions by aggregating complementary responses from average pooling and max pooling along the channel dimension. The channel attention branch and spatial attention branch are represented as follows:

[0184]

[0185]

[0186] in, Indicates channel attention characteristics; Indicates spatial attention characteristics; Indicates global average pooling; and These represent average pooling and max pooling, respectively. Represents the weights of the dimensionality-reduced convolution. ; Indicates the weights of the expanded convolution. ; To balance the channel reduction ratio and the parameter complexity of the model with the interactive learning ability of features between channels, this embodiment will... The preferred value is 16, which allows the network to capture nonlinear dependencies across channels through the bottleneck structure while avoiding excessive computational overhead. Represents the ReLU activation function; Indicates that the convolution kernel is 2D convolution; This indicates a splicing operation along the channel dimension.

[0187] The splicing and reshaping unit is used to combine channel attention features, spatial attention features, and initial aggregated features. The concatenated features are then reshaped using a tensor reshaping operation. The reshaped features are then grouped and convolved, and finally, content-adaptive spatial weights are generated using a sigmoid activation function. .

[0188] In this embodiment, the element-wise addition of channel attention features and spatial attention features serves as a structural prior; this combined prior... With initial aggregation features splicing. The spliced ​​features are then reshaped using tensor operations. The features are then rearranged. To control the model's parameter complexity while extracting local features, the reshaped features are passed through standard grouped convolutions. Processing is used to predict the importance at the location level, ultimately resulting in content-adaptive spatial weights. Represented as:

[0189]

[0190] in, This indicates the process of splicing and reshaping units.

[0191] Weighted fusion unit, used to utilize content-adaptive spatial weights For high-level features and low-level features Perform weighted fusion to combine the initial aggregated features The features are connected to the weighted fused feature residuals, and the channel dimension of the residual-connected features is optimized by pointwise convolution to output a fused feature map.

[0192] In this embodiment, content-adaptive spatial weights The complementary contributions of each feature level are strictly determined. To ensure the stability of gradient propagation and promote feature alignment, the initial aggregated features... By combining residual connections with adaptive gated features, the channel dimension is optimized through pointwise convolution.

[0193]

[0194] in, This represents the fused feature map. This unit utilizes low-level features... Higher weights are assigned to detail-sensitive regions to enhance boundary and edge structures; meanwhile, features from higher-level features are preserved in sparsely textured background regions. The semantic discriminative activations are then combined with the initial aggregated features. Seamless integration.

[0195] The fusion feature map output by the first triple attention fusion module is used as the second fusion feature map; the fusion feature map output by the second triple attention fusion module is used as the third fusion feature map.

[0196] To further eliminate background clutter and enhance target saliency before the feature input detection head, this embodiment deploys two "triple attention fusion modules" at specific nodes in the neck network of the target detection model. These two modules are responsible for the final cross-channel and spatial dimension adaptive calibration of the feature tensors to be fed into the corresponding scale detection head, thereby sending the fused, more accurate infrared features into the detection head. Specifically, the "triple attention fusion module" is responsible for deep nonlinear fusion of the low-frequency global thermal radiation background information extracted from the backbone network with the high-level semantic features returned from the neck network. This design ensures that the feature map sent to the bottom-level detection head has both a clear target semantic boundary and retains the original thermal radiation background information of the infrared weak target to the greatest extent, thereby significantly reducing the false negative rate of small targets. Infrared images are mainly composed of slowly varying low-frequency background thermal radiation and abruptly changing high-frequency local thermal radiation (i.e., targets, edges, and clutter). Due to the lack of texture and shape, infrared weak targets usually appear in images as abrupt responses of local high-frequency thermal radiation. Therefore, shallow features are crucial for capturing such high-frequency target signals; while deep features, by extracting low-frequency global thermal radiation background information, help the network understand scene context and suppress complex background clutter. Through a content-guided triple attention fusion module, the fused features are collaboratively calibrated across channel, spatial, and pixel dimensions, and fine-tuned across multiple dimensions to enhance the discriminative power of the features.

[0197] The detection head is used to perform regression and classification prediction on the multi-scale fused feature map in parallel, and to perform non-maximum suppression processing on the obtained prediction box, target category and confidence level to obtain the final infrared target detection result.

[0198] The detection head includes:

[0199] The first detection branch is used to perform regression and classification prediction on the first scale fused feature map and output the first infrared target detection result.

[0200] The second detection branch is used to perform regression and classification prediction on the second-scale fused feature map and output the second infrared target detection result.

[0201] The third detection branch is used to perform regression and classification prediction on the third-scale fused feature map and output the third infrared target detection result.

[0202] The non-maximum suppression unit is used to perform non-maximum suppression processing on the prediction boxes, target categories and confidence levels in the first prediction result, the second prediction result and the third prediction result, and output the final infrared target detection result.

[0203] In this embodiment, as Figure 2As shown, the detection head employs a multi-scale decoupled detection architecture, specifically comprising three parallel detection output branches, corresponding to large, medium, and small receptive field scales of feature maps, respectively. This aims to achieve accurate regression and classification prediction of infrared targets of different sizes (especially weak and multi-scale targets) in complex scenes. During the inference phase, to obtain the final detection result, the network concatenates all predicted boxes output from these three branches and then uniformly inputs them into a non-maximum suppression unit (NMU). The NMU filters out redundant predicted boxes for the same target based on the set confidence threshold and intersection-over-union (IoU) threshold, thereby achieving the final fusion of cross-scale detection results. The final output infrared target detection result is a structured array that includes not only the detection boxes but also the target category and confidence level.

[0204] In this embodiment, the hyperparameters used for training are set as follows: batch size is set to 8, training epochs are set to 500, the optimizer is AdamW, the initial learning rate is set to 0.001, and the warm-up epochs are set to 3. Furthermore, the size of the UAV infrared images input to the target detection model is uniformly adjusted to 640×640 pixels. During training, the preprocessed infrared image sequence is sequentially input into the backbone network, neck network, and detection head for forward propagation. The loss value between the target detection result and the true label is calculated, and the loss function includes classification loss, regression loss, and confidence loss. The gradient is calculated through backpropagation, and the network parameters are updated using the AdamW optimizer. After each training epoch, the model performance is evaluated on the validation set, and the weights of the model with the best validation performance are saved as the final target detection model.

[0205] The original infrared image sequence to be detected is acquired and preprocessed. The preprocessed infrared image sequence is then input into a trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection result.

[0206] Example 2:

[0207] This embodiment presents a UAV infrared target detection system based on dynamic feature collaboration and state space transformation, such as... Figure 5 As shown, the system includes:

[0208] The training set construction module is used to continuously acquire raw infrared images containing the target region and their corresponding ground truth labels. All acquired raw infrared images are preprocessed and associated with their corresponding ground truth labels to construct a training set.

[0209] The detection model construction module is used to build and train an infrared target detection model based on dynamic feature collaboration and state space to obtain a trained infrared target detection model based on dynamic feature collaboration and state space.

[0210] The infrared image acquisition module is used to acquire the original infrared image sequence to be detected and to preprocess it to obtain the preprocessed infrared image sequence.

[0211] The infrared target detection module is used to input the preprocessed infrared image sequence into the trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection results.

[0212] Example 3:

[0213] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the infrared target detection method based on dynamic feature collaboration and state space.

[0214] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the infrared target detection method based on dynamic feature coordination and state space as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0215] The processor is used to execute all or part of the steps in the infrared target detection method based on dynamic feature cooperation and state space as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0216] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the infrared target detection method based on dynamic feature coordination and state space described in the above embodiments.

[0217] Example 4:

[0218] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0219] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the infrared target detection method based on dynamic feature collaboration and state space described in the various embodiments of this application.

[0220] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned infrared target detection method based on dynamic feature collaboration and state space.

[0221] Example 5:

[0222] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the infrared target detection method based on dynamic feature collaboration and state space.

[0223] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0224] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0225] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. An infrared target detection method based on dynamic feature collaboration and state space, characterized in that, This method includes the following steps: Continuously acquire raw infrared images containing the target region and their corresponding real labels. Preprocess all acquired raw infrared images and associate them with their corresponding real labels to construct a training set. An infrared target detection model based on dynamic feature collaboration and state space is constructed and trained using a training set to obtain a trained infrared target detection model based on dynamic feature collaboration and state space. The original infrared image sequence to be detected is acquired and preprocessed. The preprocessed infrared image sequence is then input into a trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection result.

2. The infrared target detection method based on dynamic feature collaboration and state space according to claim 1, characterized in that, The specific method for preprocessing continuously acquired raw infrared images containing the target region is as follows: Perform the following operations frame-by-frame on the continuously acquired raw infrared images containing the target region: Blind pixel replacement and non-uniformity correction are performed on the original infrared image of the current frame to smooth fixed pattern noise and obtain the corrected infrared image. Adaptive histogram equalization or automatic gain control algorithm is used to perform dynamic range compression and linear stretching on the corrected infrared image to obtain the compressed infrared image. The compressed infrared image is subjected to channel expansion and size normalization to obtain a preprocessed infrared image, which in turn yields a preprocessed infrared image sequence.

3. The infrared target detection method based on dynamic feature collaboration and state space according to claim 2, characterized in that, The infrared target detection model based on dynamic feature collaboration and state space includes: a backbone network, a neck network, and a detection head; The backbone network is used to extract features from the preprocessed infrared image and output a multi-scale feature map. The neck network is used to perform dynamic feature collaboration and state space transformation on multi-scale feature maps and output multi-scale fused feature maps. The detection head is used to perform regression and classification prediction on the multi-scale fused feature map in parallel, and to perform non-maximum suppression processing on the obtained prediction box, target category and confidence score to obtain the final infrared target detection result. The infrared target detection results include: detection bounding box, target category, and confidence level.

4. The infrared target detection method based on dynamic feature collaboration and state space according to claim 3, characterized in that, The backbone network comprises, in sequence: a convolutional block, a first convolutional layer, a first cross-stage parallel convolutional module, a second convolutional layer, a second cross-stage parallel convolutional module, a third convolutional layer, a first dynamic calibration region attention module, a fourth convolutional layer, and a second dynamic calibration region attention module. The convolutional block is used to sequentially perform convolution, batch normalization, and SiLU activation function operations on the input preprocessed infrared image to obtain an initial feature map; The first, second, third, and fourth convolutional layers are all independent convolutional layers; The first and second cross-stage parallel convolutional modules have the same network architecture, both including: a standard convolutional layer, a channel splitting layer, a backbone computation branch, a cross-stage direct connection branch, a channel concatenation layer, and an output layer. The standard convolutional layer maps the input feature map to a preset channel dimension. The channel splitting layer equally divides the mapped feature map into a first branch feature map and a second branch feature map along the channel dimension. The backbone computation branch sequentially performs standard convolution, custom receptive field convolution, and deconvolution on the first branch feature map to obtain a deep feature map. The cross-stage direct connection branch directly passes the second branch feature map to the channel concatenation layer. The channel concatenation layer concatenates the deep feature map and the second branch feature map along the channel dimension. The output layer performs channel fusion and dimensionality reduction on the concatenated feature map to output a cross-stage fused feature map. The output of the first cross-stage parallel convolution module is the first cross-stage fused feature map; the output of the second cross-stage parallel convolution module is the second cross-stage fused feature map. The first and second dynamic calibration region attention modules have the same network architecture, both including: a dynamic calibration unit for performing lightweight dynamic calibration on the input feature map through dynamic hyperbolic tangent operation; and a dense aggregation unit for performing multi-level feature extraction and fusion on the calibrated feature map to obtain a feature representation. Residual connection units are used for feature representation. After channel-by-channel scaling, the refined feature map is added to the input feature map to obtain the refined feature map. The output of the first dynamic calibration region attention module is a first refined feature map; the output of the second dynamic calibration region attention module is a second refined feature map. The first cross-stage fusion feature map, the second cross-stage fusion feature map, the first refined feature map, and the second refined feature map are used as multi-scale feature maps.

5. The infrared target detection method based on dynamic feature collaboration and state space according to claim 4, characterized in that, The neck network includes: The attention scale sequence fusion module is used to perform dimensional matching and fusion of multi-scale feature maps to obtain a scale fusion feature map; it performs two-layer progressive feature enhancement and state space modeling on the multi-scale feature maps to obtain a first-level state feature map and a second-level state feature map; and then uses a one-dimensional convolutional channel attention mechanism to align and weight the scale fusion feature map and the second-level state feature map to obtain a first fusion feature map. The first splicing layer is used to splice the first-level state feature map and the second-level state feature map to obtain the first spliced ​​feature map; The first visual state space module is used to perform visual state space modeling on the first stitched feature map to obtain the global state feature map. The first triple attention fusion module is used to adaptively weight the first fused feature map with the global state feature map in the channel dimension, spatial dimension and pixel dimension to obtain the second fused feature map. The second concatenation layer is used to concatenate the second refined feature map in the multi-scale feature map with the global state feature map to obtain the second concatenated feature map. The third cross-stage parallel convolution module is used to perform channel mapping on the second concatenated feature map and equally divide it into two branch feature maps. The first branch feature map is then restored by standard convolution, custom receptive field convolution, and deconvolution. After being concatenated with the second branch feature map, it is fused by channels and reduced in dimension to obtain the third cross-stage fused feature map. The second and third attention fusion module is used to adaptively weight the second refined feature map and the third cross-stage fusion feature map in the multi-scale feature map in the channel dimension, spatial dimension and pixel dimension to obtain the third fusion feature map. The first fusion feature map, the second fusion feature map, and the third fusion feature map are used as multi-scale fusion feature maps.

6. The infrared target detection method based on dynamic feature collaboration and state space according to claim 5, characterized in that, The attention scale sequence fusion module includes: a scale sequence fusion module, a first triple feature enhancement module, a second visual state space module, a second triple feature enhancement module, a third visual state space module, and an attention guidance calibration module; The scale sequence fusion module is used to project the second cross-stage fused feature map, the first refined feature map, and the second refined feature map in the multi-scale feature map to the same resolution and stack them along the channel dimension. The stacked feature map is then subjected to three-dimensional batch normalization, nonlinear activation, and three-dimensional max pooling in sequence to obtain the scale fused feature map. The first triple feature enhancement module is used to enhance the second cross-stage fusion feature map, the first refined feature map and the second refined feature map in the multi-scale feature map respectively, and to stitch the enhanced feature maps along the channel dimension to obtain the first-level enhanced feature map. The second visual state space module is used to perform visual state space modeling on the first-level enhanced feature map to obtain the first-level state feature map. The second and third feature enhancement modules are used to enhance the first cross-stage fusion feature map, the second cross-stage fusion feature map, and the first-level state feature map in the multi-scale feature map respectively, and to stitch the enhanced feature maps along the channel dimension to obtain the second-level enhanced feature map. The third visual state space module is used to perform visual state space modeling on the secondary enhanced feature map to obtain the secondary state feature map. The attention-guided calibration module performs channel calibration on the secondary state feature map using a one-dimensional convolutional channel attention mechanism, and then sums it pixel-by-pixel with the scale fusion feature map to obtain residual fusion features; then, it uses a position attention component to perform spatial attention weighting on the residual fusion features to obtain the first fusion feature map.

7. The infrared target detection method based on dynamic feature collaboration and state space according to claim 6, characterized in that, The first triple feature enhancement module and the second triple feature enhancement module have the same network architecture, both of which include: defining the input three feature maps as high-resolution feature maps, medium-resolution feature maps and low-resolution feature maps in order of resolution from high to low; Adaptive max pooling and average pooling are performed simultaneously on the high-resolution feature map, and the resulting max pooled features and average pooled features are added element by element to obtain the high-resolution enhanced feature map. Nearest neighbor interpolation is performed on the medium-resolution feature map and the low-resolution feature map respectively to obtain the medium-resolution enhanced feature map and the low-resolution enhanced feature map; The obtained high-resolution enhanced feature map, medium-resolution enhanced feature map and low-resolution enhanced feature map are concatenated along the channel dimension to obtain the output enhanced feature map; The enhanced feature map output by the first triple feature enhancement module is a first-level enhanced feature map; the enhanced feature map output by the second triple feature enhancement module is a second-level enhanced feature map.

8. The infrared target detection method based on dynamic feature collaboration and state space according to claim 7, characterized in that, The first visual state space module, the second visual state space module, and the third visual state space module have the same network architecture, all including: The layer normalization unit is used to perform layer normalization on the input feature map to obtain the normalized feature map; The first depthwise convolutional unit is used to perform depthwise separable convolution on the normalized feature map to obtain a locally enhanced feature map. The hidden state mixer's state space layer is used to capture long-range dependencies in the local augmented feature map using a discrete state space model, and outputs a state augmented feature map. The feature projection layer is used to perform feature projection and batch normalization on the state-enhanced feature map, and to project and normalize the feature map. The second depthwise convolutional unit is used to perform depthwise separable convolution on the projection normalized feature map to obtain a depth-enhanced feature map. The feedforward network unit is used to perform inter-channel information exchange on the deep augmented feature map and output the channel interaction feature map. The hierarchical scaling unit is used to multiply the channel interaction feature map with a learnable diagonal weight matrix to obtain a scaled feature map. The residual connection unit is used to add the scaled feature map to the input feature map element by element to obtain the state feature map; The state feature map output by the first visual state space module is a global state feature map; the state feature map output by the second visual state space module is a first-level state feature map; and the state feature map output by the third visual state space module is a second-level state feature map.

9. The infrared target detection method based on dynamic feature collaboration and state space according to claim 8, characterized in that, The first and second triple attention fusion modules have the same network architecture, both including: defining the lower-resolution feature map from the two input feature maps as a high-level feature map. High-resolution feature maps are defined as low-level features. ; Initial aggregation unit, used to aggregate high-level features and low-level features Initial fusion is performed through element-wise addition to obtain the initial aggregated features. ; Channel attention branch, used for initial aggregated features Calculate channel attention to obtain channel attention features; Spatial attention branch, used for initial aggregated features Calculate spatial attention to obtain spatial attention features; The splicing and reshaping unit is used to combine channel attention features, spatial attention features, and initial aggregated features. The concatenated features are then reshaped using a tensor reshaping operation. The reshaped features are then grouped and convolved, and finally, content-adaptive spatial weights are generated using a sigmoid activation function. ; Weighted fusion unit, used to utilize content-adaptive spatial weights For high-level features and low-level features Perform weighted fusion to combine the initial aggregated features The features are connected to the weighted fused feature residuals, and the channel dimension of the residual-connected features is optimized by pointwise convolution to output a fused feature map. The fusion feature map output by the first triple attention fusion module is used as the second fusion feature map; the fusion feature map output by the second triple attention fusion module is used as the third fusion feature map.

10. An infrared target detection system based on dynamic feature collaboration and state space, used to implement the infrared target detection method based on dynamic feature collaboration and state space as described in any one of claims 1-9, characterized in that, The system includes: The training set construction module is used to continuously acquire raw infrared images containing the target region and their corresponding real labels, preprocess all acquired raw infrared images and associate them with their corresponding real labels to construct a training set. The detection model building module is used to build and train an infrared target detection model based on dynamic feature collaboration and state space to obtain a trained infrared target detection model based on dynamic feature collaboration and state space. The infrared image acquisition module is used to acquire the original infrared image sequence to be detected and to preprocess it to obtain the preprocessed infrared image sequence. The infrared target detection module is used to input the preprocessed infrared image sequence into the trained infrared target detection model based on dynamic feature collaboration and state space to obtain the infrared target detection results.

Images