An infrared target detection method combining multi-directional contrast attention information
By combining multi-directional contrast attention information in infrared target detection, and utilizing multi-layer feature extraction and attention information fusion, the accuracy and robustness issues in infrared small target detection are solved, achieving efficient infrared small target detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2024-03-19
- Publication Date
- 2026-07-31
AI Technical Summary
Existing infrared small target detection methods are insufficient in terms of detection accuracy and robustness, especially in complex scenarios where false alarms and missed detections are prone to occur, and there is a lack of effective utilization of feature information.
An infrared target detection method combining multi-directional contrast attention information is proposed. By using a multi-layer feature extraction network, a local Gaussian structure information extraction network, a multi-directional contrast attention information extraction network, and a multi-feature fusion network, local and global attention information is fused to achieve accurate detection of small infrared targets.
It improves the accuracy and robustness of infrared small target detection, reduces false alarm rate and false negative rate, and enhances detection performance in complex scenarios.
Smart Images
Figure CN118365923B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infrared target detection technology, and more specifically to an infrared target detection method that combines multi-directional contrast attention information. Background Technology
[0002] Infrared target detection and tracking is an important research topic with applications in military, transportation, and medical fields. In the military field, traditional radar detection systems are easily exposed and vulnerable to missile attacks; while infrared detection systems offer numerous advantages, such as passive detection, lack of strong radiation, and resistance to interference or exposure. Infrared target detection can save on hardware costs. Infrared search and track (IRST) systems have high military value, and infrared weak target detection technology is crucial in infrared search, infrared tracking, and infrared early warning. By improving target detection and tracking algorithms while maintaining the overall system integrity, we can enhance target detection speed and accuracy, giving weapon and guidance systems sufficient time for effective defense.
[0003] Current infrared small target detection methods can be broadly classified into three categories: traditional algorithms based on background suppression or visual saliency, data structure-based algorithms, and deep learning-based algorithms.
[0004] Traditional algorithms based on background suppression or visual saliency mainly analyze the background of infrared images, predicting the background component from the infrared image and then subtracting the background from the original image, or analyzing the difference between the target region and the background to design algorithms to enhance the target in the image and suppress the background. These methods include top-hat operations, maximum median filters, local contrast algorithms, multi-scale block contrast measurements, etc.
[0005] Data structure-based algorithms assume that background image patches in infrared small target images exhibit a linear correlation trend, satisfying the low-rank assumption, while the target occupies only a small part of the entire image, satisfying the sparsity assumption. Therefore, the infrared small target detection problem can be transformed into a mathematical optimization problem of recovering low-rank and sparse components from the original image. Such methods include infrared block image models, non-negative infrared block image models, rank approximation minimization models, reweighted infrared block tensor models based on nonlocal and local priors, tensor kernel norm part sum models, etc.
[0006] Deep learning-based algorithms typically use classic classification networks as the backbone for downsampling and feature extraction during the encoding stage. Between encoding and decoding, they employ summation or concatenation methods for feature fusion, and combine spatial attention or channel attention mechanisms to more effectively extract target features. These methods include asymmetric context modulation methods for infrared small target detection, attentional local contrastive networks for infrared small target detection, and internal attention-based perceptual networks for infrared small target detection. These methods utilize design strategies such as contextual feature fusion, multi-scale feature fusion, and integration with traditional methods to improve the accuracy of infrared small target detection. However, infrared small target images have limited available information, lacking texture, shape, and other feature information, which restricts the performance of infrared small target detection algorithms. Summary of the Invention
[0007] To address the aforementioned shortcomings in the prior art, this invention provides an infrared target detection method that combines multi-directional contrast attention information. By utilizing the local structural information of small infrared targets and multi-directional contrast attention information, accurate detection of small infrared targets can be achieved.
[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:
[0009] An infrared target detection method incorporating multi-directional contrast attention information includes the following steps:
[0010] Acquire infrared target images;
[0011] The infrared target image is input into a pre-trained neural network model; the neural network model includes a multi-layer feature extraction network, a local Gaussian structure information extraction network, a multi-directional contrast attention information extraction network, and a multi-feature fusion network;
[0012] A multi-scale first feature map is extracted from infrared target images using a multi-layer feature extraction network.
[0013] The local structure information is calculated for the first feature map at each scale using a local Gaussian structure information extraction network. The local structure information is then fused with the upsampled feature map of the first feature map at each scale to obtain a multi-scale second feature map.
[0014] A multi-directional contrast attention information extraction network is used to extract spatial attention information from the second feature maps at each scale, resulting in a multi-scale third feature map.
[0015] A multi-feature fusion network is used to fuse features from the third feature map at multiple scales to obtain a detection confidence map;
[0016] Adaptive threshold segmentation is applied to the detection confidence map to obtain the infrared target detection results.
[0017] Optionally, extracting a multi-scale first feature map from an infrared target image using a multi-layer feature extraction network includes:
[0018] By using a multi-layer feature extraction network, infrared target images are downsampled into feature maps of different scales of the original image size, resulting in a multi-scale first feature map.
[0019] Optionally, a local Gaussian structure information extraction network is used to calculate local structure information for the first feature maps at each scale, and the local structure information is fused with the upsampled feature maps of the first feature maps at each scale to obtain a multi-scale second feature map, including:
[0020] Gaussian convolution kernels are obtained by fitting the weights of a convolution kernel of a set size using a Gaussian function;
[0021] Gaussian convolution kernels are used to filter the first feature maps at each scale to obtain local Gaussian structure feature maps;
[0022] The first feature maps at each scale are upsampled, and the local Gaussian structure feature maps are concatenated with the upsampled feature maps. Then, convolution, batch normalization, and activation operations are performed to obtain the second feature maps at multiple scales.
[0023] Alternatively, the parameters of the Gaussian function can be updated using the backpropagation algorithm when training the neural network model.
[0024] Optionally, a multi-directional contrast attention information extraction network is used to extract spatial attention information from the second feature maps at each scale, resulting in a multi-scale third feature map, including:
[0025] Convolution operations are performed on the second feature maps at each scale to generate the corresponding feature maps themselves, related feature maps, and global feature maps.
[0026] Convolution operations are performed on the relevant feature maps using convolution kernels in four directions, and a Hadamard product is performed with the feature map itself. The four generated feature maps are then concatenated in the channel direction to obtain the concatenated feature map.
[0027] Extract a global attention feature map from the global feature map, perform a convolution operation on the concatenated feature map, perform a Hadamard product with the original feature map, and then add it to the global attention feature map to obtain a multi-scale third feature map.
[0028] Optionally, a multi-feature fusion network is used to fuse features from the multi-scale third feature map to obtain a detection confidence map, including:
[0029] A multi-feature fusion network is used to perform bilinear upsampling of the third feature map at multiple scales to obtain feature values with the same size as the original image. Then, all feature maps of the same size are stitched together. Finally, a convolution operation is performed on the stitched feature map to obtain the detection confidence map.
[0030] Alternatively, adaptive data augmentation can be performed on the infrared target image when training the neural network model.
[0031] Optionally, adaptive data augmentation of infrared target images includes:
[0032] Two infrared target images are randomly acquired;
[0033] Adaptive enhancement of the target region in another infrared target image is performed based on the gray intensity and area ratio information of the target region in one infrared target image.
[0034] The adaptively enhanced target region is pasted onto the randomly folded infrared target image to obtain the amplified infrared target image.
[0035] Alternatively, the adaptively augmented infrared target image can be preprocessed when training the neural network model.
[0036] Optionally, preprocessing the adaptively augmented infrared target image includes:
[0037] Adjust the image size of the infrared target image to the set size;
[0038] The adjusted infrared target image is then normalized.
[0039] All infrared target images are randomly divided into training and test sets according to a set ratio.
[0040] The present invention has the following beneficial effects:
[0041] This invention addresses the challenges of small target detection in infrared imaging, such as small target size, lack of texture and structural information, and imbalanced positive and negative samples. It proposes a method that combines multi-directional contrast attention information. By incorporating local structural information into a deep neural network based on the Gaussian two-dimensional distribution assumption, and by introducing a multi-directional contrast attention module, it combines local contrast attention information with global attention information. This overcomes the difficulty in detecting small targets in infrared images due to their small size, significantly improving the accuracy and robustness of infrared small target detection in complex scenes, effectively reducing false alarms, and lowering the false negative rate for small targets. Attached Figure Description
[0042] Figure 1This is a flowchart illustrating an infrared target detection method that incorporates multi-directional contrast attention information.
[0043] Figure 2 This is a schematic diagram illustrating the principle of an infrared target detection method that combines multi-directional contrast attention information.
[0044] Figure 3 A schematic diagram of the acquired infrared target image;
[0045] Figure 4 This is a schematic diagram of a network for extracting local Gaussian structure information.
[0046] Figure 5 A schematic diagram of a multi-directional contrast attention information extraction network;
[0047] Figure 6 A schematic diagram of a sample amplification method for infrared target images;
[0048] Figure 7 This is a schematic diagram of the detection results of an infrared target image. Detailed Implementation
[0049] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0050] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides an infrared target detection method that combines multi-directional contrast attention information, including the following steps S1 to S7:
[0051] S1. Acquire infrared target image;
[0052] S2. Input the infrared target image into a pre-trained neural network model; the neural network model includes a multi-layer feature extraction network, a local Gaussian structure information extraction network, a multi-directional contrast attention information extraction network, and a multi-feature fusion network;
[0053] S3. Use a multi-layer feature extraction network to extract multi-scale first feature maps from infrared target images;
[0054] S4. Calculate local structure information for the first feature map at each scale using the local Gaussian structure information extraction network, and fuse the local structure information with the upsampled feature map of the first feature map at each scale to obtain the second feature map at multiple scales.
[0055] S5. Using a multi-directional contrast attention information extraction network, spatial attention information is extracted from the second feature maps at each scale to obtain multi-scale third feature maps.
[0056] S6. Use a multi-feature fusion network to fuse features from the multi-scale third feature map to obtain the detection confidence map;
[0057] S7. Perform adaptive threshold segmentation on the detection confidence map to obtain the infrared target detection results.
[0058] This invention addresses the challenges of small target detection in infrared imaging, such as small target size, lack of texture and structural information, and imbalanced positive and negative samples. It proposes a method that combines multi-directional contrast attention information. By incorporating local structural information into a deep neural network based on the Gaussian two-dimensional distribution assumption, and by introducing a multi-directional contrast attention module, it combines local contrast attention information with global attention information. This overcomes the difficulty in detecting small targets in infrared images due to their small size, significantly improving the accuracy and robustness of infrared small target detection in complex scenes, effectively reducing false alarms, and lowering the false negative rate for small targets.
[0059] This invention combines the advantages of manual feature modeling methods and deep learning methods, making full use of the local structural information and multi-directional contrast attention information of small infrared targets to achieve accurate detection of small infrared targets. Furthermore, it employs a multi-layer information fusion method to balance the fine-grained information and semantic information of the feature maps.
[0060] In an optional embodiment of the present invention, this embodiment utilizes a multi-layer feature extraction network to extract a first feature map of multiple scales from an infrared target image, including:
[0061] By using a multi-layer feature extraction network, infrared target images are downsampled into feature maps of different scales of the original image size, resulting in a multi-scale first feature map.
[0062] For example, in this embodiment, infrared target image data is input into a multi-layer feature extraction network, and the input image is as follows: Figure 3 As shown, the image is downsampled to 1 / 3 of the original image. 2 ,1 / 2 2 1 / 4 2 1 / 8 2 1 / 16 2 The five first feature maps C i (i∈[1,5]).
[0063] In an optional embodiment of the present invention, this embodiment utilizes a local Gaussian structure information extraction network to calculate local structure information for the first feature maps at each scale, and fuses the local structure information with the upsampled feature maps of the first feature maps at each scale to obtain a multi-scale second feature map, including:
[0064] Gaussian convolution kernels are obtained by fitting the weights of a convolution kernel of a set size using a Gaussian function;
[0065] Gaussian convolution kernels are used to filter the first feature maps at each scale to obtain local Gaussian structure feature maps;
[0066] The first feature maps at each scale are upsampled, and the local Gaussian structure feature maps are concatenated with the upsampled feature maps. Then, convolution, batch normalization, and activation operations are performed to obtain the second feature maps at multiple scales.
[0067] For example, the local Gaussian structure information extraction network used in this embodiment is as follows: Figure 4 As shown, firstly, the local structural information of the same layer is calculated layer by layer. Then, the calculated local structural information of the same layer is fused with the information obtained from upsampling to obtain the original image. 2 ,1 / 2 2 1 / 4 2 1 / 8 2 1 / 16 2 The 5 second feature maps D i (i∈[1,5]).
[0068] This embodiment analyzes feature map C layer by layer. i The specific steps for calculating local structural information are as follows: using the Gaussian function... The n×n convolution kernel weights are fitted using a Gaussian function to obtain a Gaussian convolution kernel G. Finally, the fitted Gaussian convolution kernel G is used to analyze the first feature map C. i Filtering is performed to obtain the local Gaussian structure feature map L. i , where i = 1 to 4, and n has the values of n in layers C1 to C4 as n = 11, 7, 5, 3 respectively.
[0069] Specifically, in this embodiment, the backpropagation algorithm is used to update the parameters A and σ of the Gaussian function when training the neural network model. x and v y This allows for the adjustment of filter weights, enabling traditional filtering methods and deep learning methods to complement each other. This improves the feature extraction performance of the network with only a small increase in the number of parameters, and reduces the false alarm rate of the network detection results.
[0070] In this embodiment, feature maps U1 to U4 are obtained by upsampling layer by layer. iCompared with the local Gaussian structure feature map L obtained in the previous step i The images are stitched together along the channel direction, then subjected to convolution, batch normalization, and activation function operations to halve the number of channels, yielding 1 copy of the original image. 2 ,1 / 2 2 1 / 4 2 1 / 8 2 1 / 16 2 The 5 second feature maps D i (i∈[1,5]), the expression for this module is as follows:
[0071] L i =C i *G
[0072] D i =ReLu(BatchNorm(Conv) 3×3 (Concat(L i U i ))))
[0073] Where * denotes a convolution operation, G is the Gaussian convolution kernel obtained from the previous step by fitting with variable parameters, ReLU represents the activation function operation ReLU(x) = max(0,x), BatchNorm represents the batch normalization operation, and Conv 3×3 This represents a 3×3 convolution kernel with a stride of 1 and a pooling value of 1. Concat represents the concatenation operation along the channel direction, and i = 1 to 4.
[0074] In an optional embodiment of the present invention, this embodiment utilizes a multi-directional contrast attention information extraction network to extract spatial attention information from the second feature maps at various scales, thereby obtaining a multi-scale third feature map, including:
[0075] Convolution operations are performed on the second feature maps at each scale to generate the corresponding feature maps themselves, related feature maps, and global feature maps.
[0076] Convolution operations are performed on the relevant feature maps using convolution kernels in four directions, and a Hadamard product is performed with the feature map itself. The four generated feature maps are then concatenated in the channel direction to obtain the concatenated feature map.
[0077] Extract a global attention feature map from the global feature map, perform a convolution operation on the concatenated feature map, perform a Hadamard product with the original feature map, and then add it to the global attention feature map to obtain a multi-scale third feature map.
[0078] For example, the multi-directional contrast attention information extraction network used in this embodiment is as follows: Figure 5 As shown, the second feature map D is extracted layer by layer. iSpatial attention information (i∈[1,5]) is used to obtain the third feature map M. i (i∈[1,5]).
[0079] In this embodiment, a 1×1 convolution is used to generate the feature map D itself. self Related feature map D corr and global feature map D global Its channel number D is 1 / 8 of the original image.
[0080] This embodiment will use the relevant feature map D corr Convolution is performed using four convolutional kernels m1, m2, m2, and m4 in four directions, and then combined with the feature map D. self The Hadamard product is performed, concatenating the four generated feature maps e1, e2, e3, and e4 along the channel direction. This concatenation is then followed by a 1×1 convolution and a softmax function to generate the concatenated feature map E. The expression for this module is as follows:
[0081] e i =D self ⊙(D corr *m i )
[0082] E = softmax(Conv) 1×1 (concat(e1,e2,e3,e4)))
[0083] Where ⊙ represents the Hadamard product, * represents convolution, m1, m2, m2, m4 are 1×3 custom convolution kernels in four directions: 0°, 45°, 90°, and 135°, respectively, and softmax represents the activation function operation. Conv 1×1 This represents a 1×1 convolution kernel, and Concat represents the concatenation operation in the channel direction, i = 1 to 4;
[0084] This embodiment focuses on the global feature map D. global Calculate global attention to obtain the global attention feature map M. g Perform a 3×3 convolution operation on the spliced feature map E, and then convolve it with the original feature map D. self The feature map M is obtained by performing the Hadamard product. l Adding the two together yields the third feature map M, whose expression is as follows:
[0085] M g =sigmoid(conv) 5×5 (concat(Avgpool(D global ),Maxpool(D global ))))
[0086] Ml =D self ⊙Conv 3×3 (E)
[0087] M = M l +M g
[0088] Where sigmoid represents the activation function operation sigmoid(x) = 1 / (1+e) -x Conv 5×5 This represents a 5×5 convolution kernel, Concat represents the concatenation operation along the channel direction, Avgpool represents the average pooling function, Maxpool represents the max pooling function, and ⊙ represents the Hadamard product. 3×3 This represents a 3×3 convolution kernel.
[0089] This invention designs a multi-directional contrast attention module that combines local contrast attention information with global attention information, overcoming the problem of difficult detection caused by the small size of the target in infrared small target images. At the same time, the module takes into account multi-directional contrast information, which helps to reduce the false alarm rate of the detection results.
[0090] In an optional embodiment of the present invention, this embodiment utilizes a multi-feature fusion network to perform feature fusion on a multi-scale third feature map to obtain a detection confidence map, including:
[0091] A multi-feature fusion network is used to perform bilinear upsampling of the third feature map at multiple scales to obtain feature values with the same size as the original image. Then, all feature maps of the same size are stitched together. Finally, a convolution operation is performed on the stitched feature map to obtain the detection confidence map.
[0092] For example, this embodiment fuses multiple features and performs detection. Specifically, the third feature map M from the previous step is bilinearly upsampled by a factor of 2 to a feature map of the same size as the original image and then stitched together. Then, a 1×1 convolution is used to fuse multiple feature maps to obtain a confidence map result.
[0093] This invention integrates local structural information into a deep neural network, takes into account the widely accepted assumption in academia and engineering that infrared small targets conform to a two-dimensional Gaussian distribution for modeling, and sets backpropagation parameters to adjust the weights of the filter during training, so that the advantages of traditional filtering methods and deep learning methods complement each other, thereby improving the feature extraction effect of the network with only a small increase in the number of parameters and reducing the false alarm rate of the network detection results.
[0094] In an optional embodiment of the present invention, infrared small target images are acquired and an infrared small target detection dataset is constructed when training the neural network model.
[0095] This embodiment first performs adaptive data augmentation on the collected infrared small target data, such as... Figure 6 As shown, the process of this data augmentation method is as follows: randomly acquire two target images A and B, and based on the target region T in image B... B The grayscale intensity and area ratio information of target T in image A A Perform adaptive enhancement, and then target T A Paste the image into the randomly flipped image B to obtain the amplified image C.
[0096] This embodiment then preprocesses the infrared small target image data, uniformly adjusting the image size to 256×256 using the formula. Normalize the image grayscale values, where x ij x represents the gray value of pixel (i,j) in the image. min x max These represent the minimum and maximum grayscale values, respectively.
[0097] In this embodiment, the obtained dataset is randomly divided into a training set and a test set in a 7:3 ratio.
[0098] This embodiment designs a sample amplification method suitable for small infrared targets. It considers the local gray intensity information and area ratio information of small targets to perform adaptive sample amplification, which can effectively alleviate the disadvantage of the small number of infrared small target samples.
[0099] This invention, after acquiring an infrared small target image dataset, uses a designed adaptive augmentation algorithm to perform online data augmentation on the training set before inputting it into the network. The loss function is calculated for each training iteration, and the network parameters are backpropagated until the end of the training process. Finally, the test set is input into the trained network to obtain a detection confidence map. After adaptive thresholding, the final detection result is obtained, such as... Figure 7 As shown.
[0100] This invention constructs an end-to-end detection network, which can obtain the detection map simply by inputting the image to be detected, without any additional manual parameter tuning or calculation steps, making it suitable for applications of automatic detection of small infrared targets.
[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0104] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0105] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. An infrared target detection method incorporating multi-directional contrast attention information, characterized in that, Includes the following steps: Acquire infrared target images; The infrared target image is input into a pre-trained neural network model; the neural network model includes a multi-layer feature extraction network, a local Gaussian structure information extraction network, a multi-directional contrast attention information extraction network, and a multi-feature fusion network; A multi-scale first feature map is extracted from infrared target images using a multi-layer feature extraction network. The local structure information is calculated for the first feature map at each scale using a local Gaussian structure information extraction network. The local structure information is then fused with the upsampled feature map of the first feature map at each scale to obtain a multi-scale second feature map. A multi-directional contrast attention information extraction network is used to extract spatial attention information from the second feature maps at each scale, resulting in a multi-scale third feature map. A multi-feature fusion network is used to fuse features from the third feature map at multiple scales to obtain a detection confidence map; Adaptive threshold segmentation is applied to the detection confidence map to obtain the infrared target detection results.
2. The infrared target detection method of claim 1, wherein, Extracting multi-scale first feature maps from infrared target images using a multi-layer feature extraction network includes: By using a multi-layer feature extraction network, infrared target images are downsampled into feature maps of different scales of the original image size, resulting in a multi-scale first feature map.
3. The method according to claim 1, wherein, A local Gaussian structure information extraction network is used to calculate local structure information for the first feature maps at each scale. This local structure information is then fused with the upsampled feature maps of the first feature maps at each scale to obtain a multi-scale second feature map, including: Gaussian convolution kernels are obtained by fitting the weights of a convolution kernel of a set size using a Gaussian function; Gaussian convolution kernels are used to filter the first feature maps at each scale to obtain local Gaussian structure feature maps; The first feature maps at each scale are upsampled, and the local Gaussian structure feature maps are concatenated with the upsampled feature maps. Then, convolution, batch normalization, and activation operations are performed to obtain the second feature maps at multiple scales.
4. The infrared target detection method of claim 3, wherein, When training the neural network model, the backpropagation algorithm is used to update the parameters of the Gaussian function.
5. The method according to claim 1, wherein, A multi-directional contrast attention information extraction network is used to extract spatial attention information from the second feature maps at each scale, resulting in a multi-scale third feature map, including: Convolution operations are performed on the second feature maps at each scale to generate the corresponding feature maps themselves, related feature maps, and global feature maps. Convolution operations are performed on the relevant feature maps using convolution kernels in four directions, and a Hadamard product is performed with the feature map itself. The four generated feature maps are then concatenated in the channel direction to obtain the concatenated feature map. Extract a global attention feature map from the global feature map, perform a convolution operation on the concatenated feature map, perform a Hadamard product with the original feature map, and then add it to the global attention feature map to obtain a multi-scale third feature map.
6. The method of claim 1, wherein, A multi-feature fusion network is used to fuse features from multiple scales of the third feature map to obtain a detection confidence map, including: A multi-feature fusion network is used to perform bilinear upsampling of the third feature map at multiple scales to obtain feature values with the same size as the original image. Then, all feature maps of the same size are stitched together. Finally, a convolution operation is performed on the stitched feature map to obtain the detection confidence map.
7. The method according to claim 1, wherein, Adaptive data augmentation is performed on infrared target images during neural network model training.
8. The method according to claim 7, wherein, Adaptive data augmentation of infrared target images includes: Two infrared target images are randomly acquired; Adaptive enhancement of the target region in another infrared target image is performed based on the gray intensity and area ratio information of the target region in one infrared target image. The adaptively enhanced target region is pasted onto the randomly folded infrared target image to obtain the amplified infrared target image.
9. The method according to claim 8, wherein, When training the neural network model, the infrared target image after adaptive data augmentation is preprocessed.
10. The method according to claim 9, wherein, Preprocessing of the adaptively augmented infrared target image includes: Adjust the image size of the infrared target image to the set size; The adjusted infrared target image is then normalized. All infrared target images are randomly divided into training and test sets according to a set ratio.