An infrared sea surface target detection method based on global-local fusion attention

Abstract

The application relates to an infrared sea surface target detection method based on global-local fusion attention, which comprises the following steps: acquiring an infrared sea surface image for pretreatment; inputting the pretreated infrared sea surface image into a main network based on global-local fusion attention for feature extraction to obtain multi-scale feature information; inputting the multi-scale feature information into a self-adaptive multi-scale feature fusion encoder for intra-scale interaction and self-adaptive cross-scale fusion to obtain multi-scale fusion feature information; adopting query selection to select fusion feature information from the multi-scale fusion feature information as an initial object query, and adopting a decoder to convert the initial object query into a target detection result. Compared with the prior art, the application has the advantages of improving the accuracy and stability of infrared sea surface target detection.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to an infrared sea surface target detection method based on global-local fusion attention. Background Technology

[0002] With the increasing maritime activities and rapid development of maritime trade, maritime security has become a key international concern. Infrared sea surface target detection technology, due to its reliable nighttime detection capabilities, longer detection range, and all-weather operation, demonstrates enormous application potential in the maritime field. The introduction of this technology will significantly improve the efficiency of maritime safety management, providing strong technical support for areas such as maritime navigation safety, maritime rescue, environmental monitoring, and coastline surveillance.

[0003] Despite the numerous advantages of infrared sea surface target detection technology, it still faces many challenges in practical applications. Factors such as dense fog and sea clutter can limit the feature information, blur edges, and lack texture details in infrared target images, thus increasing the difficulty of detection. Furthermore, the vastness of the sea area leads to significant variations in the scale of targets in infrared images, especially at long distances, where many targets appear small, increasing the likelihood of false positives and false negatives.

[0004] To address the aforementioned challenges in infrared sea surface target detection, deep learning-based methods typically incorporate attention mechanisms to enhance the salient features of infrared targets while suppressing background noise to improve target information. However, existing attention mechanisms, such as Squeeze Excitation (SE) channel attention and Convolutional Block Attention Module (CBAM), suffer from limitations in accuracy and stability when applied to infrared sea surface target detection. SE channel attention utilizes global average pooling to extract channel weights from the image, strengthening target feature channels and suppressing redundant channels. However, it neglects local details of spatial location, hindering infrared target feature enhancement. CBAM enhances image feature representation from both channel and spatial dimensions; however, it faces the challenge of information alignment when processing channel and spatial attention information, thus limiting its infrared feature extraction capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an infrared sea surface target detection method based on global-local fusion attention that improves feature extraction capabilities and enables accurate target detection.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] An infrared sea surface target detection method based on global-local fusion attention includes the following steps:

[0008] Acquire infrared sea surface images and perform preprocessing;

[0009] The preprocessed infrared sea surface image is input into a backbone network based on global-local fusion attention for deep feature extraction to obtain multi-scale feature information.

[0010] The multi-scale feature information is input into an adaptive multi-scale feature fusion encoder for intra-scale interaction and adaptive cross-scale fusion to obtain multi-scale fused feature information.

[0011] The query selection method selects fused feature information from multi-scale fused feature information as the initial object query, and the decoder transforms the initial object query into the target detection result.

[0012] Furthermore, the preprocessing step includes:

[0013] The infrared sea surface image is adaptively scaled to a uniform length and width.

[0014] Furthermore, the backbone network based on global-local fusion attention includes a Stem module, a GLFA-Block module, and a DWConv module. This backbone network employs five stages for deep feature extraction to obtain multi-scale feature information. The specific steps include:

[0015] Based on the preprocessed infrared sea surface image, the Stem module is used to perform the first stage of depth feature extraction and output the feature information of the first stage.

[0016] Based on the feature information of the first stage, the GLFA-Block module is used to extract the deep features in the second stage and output the feature information of the second stage.

[0017] Based on the feature information of the second stage, the DWConv module and GLFA-Block module are used sequentially to extract the deep features of the third stage and output the feature information of the third stage.

[0018] Based on the feature information of the third stage, the DWConv module and two GLFA-Block modules are used sequentially to extract the depth features of the fourth stage, and the feature information of the fourth stage is output.

[0019] Based on the feature information from the fourth stage, the DWConv module and the GLFA-Block module are used sequentially to extract the deep features in the fifth stage, and the feature information of the fifth stage is output.

[0020] The feature information of the third stage, the feature information of the fourth stage, and the feature information of the fifth stage are used as the multi-scale feature information.

[0021] Furthermore, the GLFA-Block module is a residual network structure, and the execution steps of each GLFA-Block module include:

[0022] The input feature information is subjected to multiple consecutive convolutional residual operations, and the operation results are merged by channel to obtain feature channel information;

[0023] Feature channel information is compressed through multiple convolution operations, and the GLFA submodule is used to simultaneously acquire global and local feature information and output it.

[0024] Furthermore, the GLFA submodule includes a global-local information extraction unit, an information fusion unit, and a multilayer perceptron. The execution steps of the GLFA submodule include:

[0025] Based on the global-local information extraction unit, global feature information is extracted using adaptive global average pooling, and local feature information is extracted using adaptive local average pooling operations at different scales. The expressions for the adaptive global average pooling operation and the adaptive local average pooling operation are as follows:

[0026] G = P(X, 1)

[0027] L1 = Conv(P(X,2))

[0028] L2 = Conv(P(X,4))

[0029] In the formula, G represents global feature information, P(·,·) represents adaptive average pooling operation, the number 1 indicates that the two-dimensional image features are transformed into one-dimensional data, X represents the input image features, L1 and L2 represent local feature information, Conv(·) represents convolution operation, and the numbers 2 and 4 represent that the feature information is transformed into dimensions of 2×2 and 4×4, respectively.

[0030] Based on the information fusion unit, the global feature information and local feature information are converted into one-dimensional channel data, and stacking and scaling are performed to fuse the global feature information and local feature information, resulting in fused global and local feature information. The expression of the fused global and local feature information is as follows:

[0031] T=Cat(Res(G),Res(L1),Res(L2))

[0032] In the formula, T represents the fused global and local feature information, Res(·) represents the feature size adjustment, and Cat(·) represents the feature channel connection operation;

[0033] Based on the fused global and local feature information, the multilayer perceptron is used for processing to obtain the final global and local feature information. The expressions for the final global and local feature information are as follows:

[0034]

[0035] In the formula, S represents the final global and local feature information, FC(·) represents the fully connected operation of the fully connected layer, δ(·) represents the ReLU activation function, and σ(·) represents the Sigmoid function. This indicates multiplication by channel.

[0036] Furthermore, the adaptive multi-scale feature fusion encoder includes an IFIA module and an adaptive feature fusion module. The IFIA module performs intra-scale interaction on the input multi-scale feature information to obtain interactive multi-scale feature information. The adaptive feature fusion module performs adaptive cross-scale fusion on the interactive multi-scale feature information to obtain multi-scale fused feature information.

[0037] Furthermore, the IFIA module employs a multi-head attention mechanism to obtain multi-scale feature information of the interaction during the intra-scale interaction process. Specifically, the operation is as follows:

[0038] Q l =K l =V l =Flatten(S) l )

[0039] A l =Reshape(Attn(Q) l ,K l V l ))

[0040] In the formula, A l For the feature information of layer l after interaction, Attn(·) represents the multi-head self-attention mechanism, and Reshape(·) represents restoring the feature shape to the same shape as S. l Consistent characteristics, S l Given the input l-layer feature information, Flatten(·) is the inverse operation of Reshape(·), Q l To retrieve the content, K l V is a candidate keyword. l This is actual information.

[0041] Furthermore, the adaptive feature fusion module uses a training-based method to determine the optimal fusion strategy, adaptively fusing features at different levels. The specific fusion operation is as follows:

[0042] El =α l ·F 3→l +β l ·F 4→l +γ l ·F 5→l

[0043]

[0044] In the formula, E l α represents the fused feature information of layer l after multi-scale feature fusion; l β l and γ l The adaptive weights learned during network training adaptively fuse feature information at different scales, satisfying α. l ,β l ,γ l ∈[0,1] and α l +β l +γ l =1 condition, α l β l and γ l The weights are respectively based on and The softmax function with parameters is used to obtain the result. The parameters are obtained through backpropagation from multi-scale feature information {F3, F4, F5} via convolution operations; F n→l This represents the feature information mapped from layer n to layer l, where n, l ∈ {3, 4, 5}. The mapping operation includes upsampling and downsampling, with the aim of aligning the scale of the features.

[0045] Furthermore, the decoder consists of multiple stacked Transformer layers. Each Transformer layer associates the initial object query with multi-scale fused feature information and uses a self-attention mechanism for further processing, ultimately outputting the object detection result.

[0046] Furthermore, the target detection results include the category, bounding box coordinates, and confidence level of the infrared sea surface target.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) This invention uses a global-local fusion attention mechanism to simultaneously capture global context information and local detail information in infrared sea surface images, enhance the edge and texture details of infrared targets, improve the feature extraction capability of infrared sea surface targets, and then achieve more accurate identification and positioning of infrared sea surface targets through multi-scale fusion and decoding.

[0049] (2) This invention uses a backbone network based on global-local fusion attention to extract deep features, which can simultaneously focus on global structural information and local detail information, enhance infrared target edge features, and improve the accuracy and stability of infrared sea surface target detection. The GLFA submodule obtains multi-scale local information by using adaptive average pooling at different scales. In order to enhance the contextual detail information ignored by block-based local average pooling operations, GLFA introduces convolution operations to enhance information interaction between blocks, enhance information exchange between different local regions, and further improve the model's adaptability to complex infrared sea surface environments.

[0050] (3) The present invention can adaptively adjust the fusion weights according to the feature information of different scales through the adaptive multi-scale feature fusion encoder, so that the model can better handle infrared sea surface targets of different sizes.

[0051] (4) The present invention can improve the accuracy and stability of infrared sea surface target detection in complex marine environments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0053] Figure 2 This is a diagram of the overall network structure of the present invention;

[0054] Figure 3 This is a structural diagram of the GLFA-Block module of the present invention;

[0055] Figure 4 This is a structural diagram of the GLFA submodule of the present invention;

[0056] Figure 5 This is a structural diagram of the IFIA module of the present invention;

[0057] Figure 6 This is a comparison chart of the detection results of the present invention. Detailed Implementation

[0058] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0059] This embodiment provides an infrared sea surface target detection method based on global-local fusion attention. The method includes: extracting target features at different scales from infrared sea surface images using a backbone network based on global-local fusion attention. Global-local fusion attention consists of global-local information extraction, information fusion, and multi-layer perception. An inter-block information exchange module is added after each local information extraction. The backbone network outputs features at different scales as input to an adaptive multi-scale feature fusion encoder. The adaptive multi-scale feature fusion encoder fully fuses multi-scale infrared feature information through intra-scale interaction and adaptive cross-scale fusion. Query selection is used to select features from the encoder output as the initial object query, and the decoder transforms the initial object query into target bounding boxes and classification confidence scores. Figure 1 As shown, the specific steps of this method are as follows:

[0060] Step 1: Acquire infrared sea surface images and perform preprocessing.

[0061] Infrared sea surface images are acquired by shipborne, airborne, or shore-based infrared thermal imagers. The acquired infrared sea surface images are adaptively scaled to transform the image size to a uniform length and width.

[0062] Step 2: Input the preprocessed infrared sea surface image into the backbone network based on global-local fusion attention for feature extraction to obtain multi-scale feature information.

[0063] Infrared sea surface images are used to extract depth feature information at different levels through a backbone network based on Global Local Fusion Attention (GLFA). The backbone network, based on GLFA, mainly consists of a Stem (stem network) module, a GLFA-Block module, and a DWConv (depth-separable convolution) module, as follows: Figure 2 As shown, the backbone network consists of five stages. The output feature size of stage 1 is 1 / 4 of the original image size, while the output feature size of each of the remaining stages is 1 / 2 of the input feature size of that stage. The feature information output from stages 3, 4, and 5 is used as the input to the adaptive multi-scale feature fusion encoder.

[0064] The GLFA-Block module is a component of the backbone network for global and local fusion attention; it has a residual network structure, such as... Figure 3 As shown, the GLFA-Block module first performs six consecutive convolutional residual operations on the input features with kernel size K and stride size 1, and then merges the output features to obtain feature channel information. Secondly, it compresses the feature channel information through two convolution operations. Finally, it uses the GLFA submodule to simultaneously obtain global and local feature information to enhance the infrared target feature information.

[0065] The GLFA submodule mainly consists of a global-local information extraction unit, an information fusion unit, and a multilayer perceptron, such as Figure 4 As shown. The specific operation flow of this GLFA submodule is as follows:

[0066] (1) The global information extraction unit adopts global average pooling operation, and the specific operation is as follows:

[0067] G = P(X, 1)

[0068] Where P(·,·) is the adaptive average pooling operation, and the number 1 indicates that the two-dimensional image features are transformed into one-dimensional data. The input image features.

[0069] Multi-scale local information is obtained by using adaptive average pooling at different scales. To enhance the contextual details ignored by block-based local average pooling operations, convolution operations are introduced to enhance information interaction between blocks, as follows:

[0070] L1 = Conv(P(X,2))

[0071] L2 = Conv(P(X,4))

[0072] Here, Conv(·) represents the convolution operation, and the numbers 2 and 4 represent transforming the feature information into dimensions of 2×2 and 4×4, respectively.

[0073] (2) The information fusion unit first converts the global and local information features into one-dimensional channel data; then, after the channel is stacked and scaled, the feature information is used to achieve the fusion of global and local information, as follows:

[0074] T=Cat(Res(G),Res(L1),Res(L2))

[0075] Res(·) represents feature size adjustment, and Cat(·) represents feature channel connection operation.

[0076] (3) The fused features are processed through a multilayer perceptron to capture the complex inter-channel relationships and calculate the weight of each channel. Then, each channel of the input features is multiplied by its corresponding weight and summed to obtain the final output:

[0077]

[0078] The multilayer perceptron consists of two fully connected layers, denoted by FC(·); δ(·) represents the ReLU activation function; and σ(·) represents the Sigmoid function. This indicates multiplication by channel.

[0079] Step 3: Input the multi-scale feature information into the adaptive multi-scale feature fusion encoder to perform intra-scale interaction and adaptive cross-scale fusion to obtain multi-scale fused feature information.

[0080] The adaptive multi-scale feature fusion encoder includes an IFIA (Intra-scale Feature Interaction Attention) module and an adaptive feature fusion module. First, the IFIA module performs intra-scale interaction on the input multi-scale features {S3, S4, S5} to obtain multi-scale features {A3, A4, A5}. Then, based on the adaptive feature fusion module, it applies bottom-up, top-down, and laterally connected multi-scale feature fusion methods to obtain multi-scale features {F3, F4, F5}. Finally, adaptive cross-scale fusion is performed to obtain the output {E3, E4, E5}, as shown below. Figure 2 As shown.

[0081] Specifically, the encoder's intra-scale interaction is achieved through the IFIA module, which performs intra-scale information exchange on the input features {S3, S4, S5} at different scales to obtain multi-scale features {A3, A4, A5}. The IFIA module employs a multi-head self-attention mechanism to enrich semantic feature information and capture the relevant relationships within the features, such as... Figure 5 As shown, the specific operation is as follows:

[0082] Q l =K l =V l =Flatten(S) l )

[0083] A l =Reshape(Attn(Q) l ,K l V l ))

[0084] Where Attn(·) represents the multi-head self-attention mechanism, and Reshape(·) represents restoring the feature shape to the shape of S. l With consistent features, Flatten(·) is the inverse operation of Reshape(·), Q l K l V l These represent the query, key, and value, respectively, and are used for attention calculation between feature maps at the same scale.

[0085] The encoder's adaptive feature fusion module performs multi-scale feature fusion on the multi-scale features {F3, F4, F5} to obtain multi-scale features {E3, E4, E5}, such as... Figure 2As shown. The adaptive feature fusion module determines the optimal fusion strategy based on training methods, adaptively fusing features at different levels. The specific fusion operation is as follows:

[0086] E l =α l ·F 3→l +β l ·F 4→l +γ l ·F 5→l

[0087] Among them, E l F represents the feature information of layer l after multi-layer feature fusion; n→l This represents the feature information mapped from layer n to layer l, where n, l ∈ {3, 4, 5}; the mapping operation includes upsampling and downsampling, the purpose of which is to align the scale of the features; α l β l and γ l This represents the adaptive weights learned during network training, which adaptively fuse feature information from different levels, satisfying α. l ,β l ,γ l ∈[0,1] and α l +β l +γ l =1 condition.

[0088] α l β l and γ l The weights are respectively based on and The softmax function with parameters is used to obtain the result, where α l The weights are represented as:

[0089]

[0090] β l and γ l Indicates with α l Similarly, among them, The parameters are obtained by backpropagation from the multi-scale feature information {F3,F4,F5} through convolution operations.

[0091] Step 4: Select fused feature information from multi-scale fused feature information as the initial object query using query selection, and use a decoder to transform the initial object query into the target detection result.

[0092] The decoder consists of multiple stacked Transformer layers. Each layer first associates the initial object query with the feature map output by the encoder, and then uses a self-attention mechanism to further process the features. The multi-layered stacked design allows the decoder to progressively refine the initial object query, generating more accurate predictions. In the final layer of the decoder, the category, bounding box coordinates, and confidence score of the infrared sea surface target are generated.

[0093] In the above embodiment, step 1 is infrared sea surface image preprocessing, step 2 is infrared sea surface target multi-scale depth feature extraction, step 3 is adaptive multi-scale feature fusion, and step 4 is obtaining the infrared sea surface target category and bounding box coordinates. In the actual infrared sea surface target detection process, steps 1, 2, 3, and 4 can be sequentially repeated to achieve target detection. Furthermore, this embodiment requires pre-training of the network model to better complete the target detection process.

[0094] For network training, an infrared maritime dataset was used, randomly divided into training, validation, and test sets, accounting for 80%, 10%, and 10% of the data, respectively. Due to the limited quantity of infrared sea surface datasets, this embodiment employs data augmentation methods such as random color adjustment, random scaling, and random cropping to enrich the dataset and improve the performance of infrared sea surface target detection. For network training, the batch size was set to 16, the number of iterations to 200, the Adam optimizer was used, the base learning rate was 0.0001, and the weight decay was 0.0001.

[0095] This embodiment performs infrared sea surface target testing on a test set, and the test results are as follows: Figure 6 As shown, the left side represents the recognition result of the YOLOv8 algorithm, the middle side represents the recognition result of the YOLOv9 algorithm, and the right side represents the recognition result of this example. Compared with the YOLOv8 and YOLOv9 algorithms, this example can effectively detect infrared sea surface targets in scenarios with similar objects, occlusion, targets of different sizes, and small targets at a distance, demonstrating high accuracy.

[0096] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0098] 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.

[0099] 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.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus 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.

[0101] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0102] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An infrared sea surface target detection method based on global-local fusion attention, characterized in that, Includes the following steps: Acquire infrared sea surface images and perform preprocessing; The preprocessed infrared sea surface image is input into a global-local fusion attention-based backbone network for deep feature extraction to obtain multi-scale feature information. This backbone network includes a Stem module, a GLFA-Block module, and a DWConv module. The global-local fusion attention-based backbone network employs five stages for deep feature extraction to obtain multi-scale feature information. Specific steps include: Based on the preprocessed infrared sea surface image, the Stem module is used to perform the first stage of depth feature extraction and output the feature information of the first stage. Based on the feature information of the first stage, the GLFA-Block module is used to extract the deep features in the second stage and output the feature information of the second stage. Based on the feature information of the second stage, the DWConv module and GLFA-Block module are used sequentially to extract the deep features of the third stage and output the feature information of the third stage. Based on the feature information of the third stage, the DWConv module and two GLFA-Block modules are used sequentially to extract the depth features of the fourth stage, and the feature information of the fourth stage is output. Based on the feature information from the fourth stage, the DWConv module and the GLFA-Block module are used sequentially to extract the deep features in the fifth stage, and the feature information of the fifth stage is output. The feature information of the third stage, the feature information of the fourth stage, and the feature information of the fifth stage are used as the multi-scale feature information; The GLFA-Block module is a residual network structure, and the execution steps of each GLFA-Block module include: The input feature information is subjected to multiple consecutive convolutional residual operations, and the operation results are merged by channel to obtain feature channel information; Feature channel information is compressed through multiple convolution operations. The GLFA submodule is used to simultaneously acquire global and local feature information and output it. The GLFA submodule includes a global and local information extraction unit, an information fusion unit, and a multilayer perceptron. The multi-scale feature information is input into an adaptive multi-scale feature fusion encoder for intra-scale interaction and adaptive cross-scale fusion to obtain multi-scale fused feature information. The query selection method selects fused feature information from multi-scale fused feature information as the initial object query, and the decoder transforms the initial object query into the target detection result.

2. The infrared sea surface target detection method based on global-local fusion attention according to claim 1, characterized in that, The preprocessing steps include: The infrared sea surface image is adaptively scaled to a uniform length and width.

3. The infrared sea surface target detection method based on global-local fusion attention as described in claim 1, characterized in that, The execution steps of the GLFA submodule include: Based on the global-local information extraction unit, global feature information is extracted using adaptive global average pooling, and local feature information is extracted using adaptive local average pooling operations at different scales. The expressions for the adaptive global average pooling operation and the adaptive local average pooling operation are as follows: In the formula, G For global feature information, For adaptive average pooling, the number 1 indicates that two-dimensional image features are transformed into one-dimensional data. For the input image features, L 1. L 2 represents local feature information. This is represented as a convolution operation, where the numbers 2 and 4 indicate that the feature information is transformed into dimensions of 2×2 and 4×4, respectively. Based on the information fusion unit, the global feature information and local feature information are converted into one-dimensional channel data, and stacking and scaling are performed to fuse the global and local feature information, resulting in fused global and local feature information. The expression of the fused global and local feature information is as follows: In the formula, T The fused global and local feature information, To adjust the feature size, This is represented as a feature channel connection operation; Based on the fused global and local feature information, the multilayer perceptron is used for processing to obtain the final global and local feature information. The expressions for the final global and local feature information are as follows: In the formula, S For the final global and local feature information, This indicates the fully connected operation of the fully connected layer. Represents the ReLU activation function. This represents the Sigmoid function. This indicates multiplication by channel.

4. The infrared sea surface target detection method based on global-local fusion attention according to claim 1, characterized in that, The adaptive multi-scale feature fusion encoder includes an IFIA module and an adaptive feature fusion module. The IFIA module performs intra-scale interaction on the input multi-scale feature information to obtain interactive multi-scale feature information. The adaptive feature fusion module performs adaptive cross-scale fusion on the interactive multi-scale feature information to obtain multi-scale fused feature information.

5. The infrared sea surface target detection method based on global-local fusion attention according to claim 4, characterized in that, The IFIA module employs a multi-head attention mechanism to obtain multi-scale feature information of the interaction during the intra-scale interaction process. Specifically, the operation is as follows: In the formula, A l For interaction l Layer feature information, This indicates a multi-head self-attention mechanism. This indicates that the feature shape is restored to the same shape as... Consistent characteristics For input l Layer feature information, yes The reverse operation, For the content to be searched, Candidate keywords, This is actual information.

6. The infrared sea surface target detection method based on global-local fusion attention according to claim 4, characterized in that, The adaptive feature fusion module uses a training-based method to determine the optimal fusion strategy, adaptively fusing features at different levels. The specific fusion operation is as follows: In the formula, Indicates the fusion of multi-scale features Layer fusion feature information; , and This represents the adaptive weights learned during network training, which adaptively fuse feature information at different scales to satisfy... and condition, , and The weights are respectively based on , and The softmax function with parameters is used to obtain the result. The parameters are derived from multi-scale feature information through convolution operations. Obtained via backpropagation; express Layer feature mapping to Layer feature information, The mapping operation includes upsampling and downsampling, with the aim of aligning the scale of features.

7. The infrared sea surface target detection method based on global-local fusion attention according to claim 1, characterized in that, The decoder consists of multiple stacked Transformer layers. Each Transformer layer associates the initial object query with multi-scale fused feature information and uses a self-attention mechanism for further processing, ultimately outputting the object detection result.

8. The infrared sea surface target detection method based on global-local fusion attention according to claim 1, characterized in that, The target detection results include the category, bounding box coordinates, and confidence level of the infrared sea surface target.

Images