Infrared-radar multi-mode iceberg target detection method based on heterogeneous feature alignment and information interaction network
By fusing features from infrared and navigation radar images through the RIFANet network, the problem of limited detection capabilities of a single sensor is solved, achieving high-accuracy detection of iceberg targets and improving recognition capabilities in complex environments.
Patent Information
- Application Number
- CN202511749312.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-02-17
AI Technical Summary
In existing technologies, the limited detection capabilities of a single sensor lead to low accuracy in iceberg target detection, especially in complex environments where it is difficult to effectively identify iceberg targets.
An infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network is adopted. Feature fusion is performed through RIFANet network. By utilizing the complementary features of infrared images and navigation radar images, a cross-modal feature alignment module and a multi-stage feature fusion module are designed to improve the detection accuracy.
It improves the accuracy of iceberg target detection, can effectively identify iceberg targets in complex environments, reduces false alarms and missed detections, and enhances the robustness of detection under poor visual conditions and clutter interference.
Smart Images
Figure CN121545008A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target detection technology, and in particular to an infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network. Background Technology
[0002] The Arctic shipping route is of great significance to the global economy, trade, shipping, and energy sectors. With the continuous melting of sea ice, the navigable period in the Arctic shipping route is increasing, and shipping activity in polar waters is becoming more frequent. However, icebergs float year-round in the Arctic Ocean. During their formation, icebergs break apart into many smaller icebergs and ice fragments. These targets are often small in size and irregularly shaped, posing challenges to ship navigation and operations. Therefore, iceberg target detection methods based on shipborne detection systems can better observe the ship's surroundings, promptly and effectively detect hazards, and make decisions to avoid property damage or even casualties caused by ship-ice collisions.
[0003] Currently, target monitoring during navigation primarily relies on navigation radar, visible light cameras, or infrared cameras. Navigation radar is the most common sensor in maritime navigation. It exhibits strong stability in complex environments and can detect targets at considerable distances in advance. However, radar cannot identify target types and can only provide approximate location and speed information during navigation. Furthermore, iceberg target echoes resemble sea clutter, easily causing confusion during detection, resulting in low accuracy for radar detection of iceberg targets. Visible light cameras acquire abundant visual information, are low-cost, and simple to operate. However, they are limited by fog and lighting conditions. Arctic weather is often accompanied by dense fog and snow, with three to four months of polar night each year, limiting the capabilities of visible light cameras and leading to low accuracy in iceberg target identification. Infrared cameras acquire information by capturing the temperature difference between the target and its surroundings. Unlike visible light cameras, they are not limited by fog and lighting conditions and can operate in low-visibility conditions, providing thermal radiation and contour information of objects. Furthermore, in actual detection, the reflection of sunlight on the sea surface produces strong flares, and the undulating waves also create localized bright areas. These areas are much brighter than the target in infrared images, exhibiting target-like characteristics that are easily confused with the target, thus interfering with infrared iceberg detection and causing false alarms. In addition, the detection range of infrared cameras is limited. When the target is very far away or its own infrared radiation is low, the target and background exhibit low contrast, making it difficult to detect and leading to missed detections. Therefore, existing methods still suffer from low accuracy in iceberg target detection due to the limited detection capabilities of a single sensor. Summary of the Invention
[0004] This invention addresses the problem of low accuracy in iceberg target detection methods due to the limited detection capabilities of single sensors, and proposes an infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction networks.
[0005] The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network includes the following steps:
[0006] Acquire the infrared image and the navigation radar image to be tested, and input the infrared image and the navigation radar image to be tested into the trained RIFANet network to obtain the iceberg target detection result;
[0007] The RIFANet network includes: an encoder module, a cross-modal feature alignment module, and a multi-stage feature fusion module;
[0008] The encoder module uses infrared images and navigation radar images to obtain infrared image feature sets and navigation radar image feature sets;
[0009] The cross-modal feature alignment module uses features from the infrared image feature set and the navigation radar image feature set to offset and align the infrared image features to obtain aligned infrared image features.
[0010] The multi-stage feature fusion module uses aligned infrared image features and infrared image feature sets to obtain iceberg target detection results.
[0011] Furthermore, the trained RIFANet network is obtained in the following way:
[0012] The infrared image of the iceberg target, the navigation radar image of the iceberg target, and the corresponding iceberg target detection results are combined into a training set. The RIFANet network is trained using the training set to obtain the trained RIFANet network.
[0013] Furthermore, the encoder module includes: a detection mode branch and a reference mode branch;
[0014] The detection modality branch utilizes infrared images to acquire infrared image feature sets. ;
[0015] The detection modality branch includes: a first VGG unit, a second VGG unit, a third VGG unit, and a Transformer unit;
[0016] The input to the first VGG unit is an infrared image of the iceberg target, and the output of the first VGG unit is the infrared image features. The input to the second VGG unit is infrared image features. The output is infrared image features. The input to the third VGG unit is infrared image features. The output is infrared image features. The input to the Transformer unit is infrared image features. The output is infrared image features. ;
[0017] The reference mode branch uses navigation radar images to obtain a set of navigation radar image features. ;
[0018] The reference mode branch includes: a first navigation radar image feature acquisition unit, a second navigation radar image feature acquisition unit, a third navigation radar image feature acquisition unit, and a fourth navigation radar image feature acquisition unit;
[0019] The first navigation radar image feature acquisition unit, the second navigation radar image feature acquisition unit, the third navigation radar image feature acquisition unit, and the fourth navigation radar image feature acquisition unit each include one A convolutional layer and a max-pooling layer;
[0020] The first navigation radar image feature acquisition unit takes a navigation radar image as input and outputs navigation radar image features. The second navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. The input to the third navigation radar image feature acquisition unit is the navigation radar image feature. The output is navigation radar image features. The fourth navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. .
[0021] Furthermore, the infrared image features , , Obtained through the following methods:
[0022]
[0023] in, It is an infrared image of an iceberg target. yes convolution, It is an activation function. It is a feature icon number; when hour, , This is an infrared image of an iceberg target.
[0024] Furthermore, the cross-modal feature alignment module includes: a spatial enhancement feature acquisition unit, a shared feature enhancement unit, an offset feature acquisition unit, and an offset alignment unit;
[0025] The spatial enhancement feature acquisition unit is used to spatially enhance the infrared image features and navigation radar image features, and to acquire the spatial enhancement features of the infrared image and the navigation radar image, specifically:
[0026]
[0027]
[0028] in, It's about spatial awareness and operation. yes function, It's a link operation. It is a global max pooling operation in the spatial dimension. It is a spatial dimension average pooling operation. , It is a feature icon number. It's a multiplication of the channel dimension. It is the first infrared image Spatial enhancement features, It is the first navigation radar image Individual spatial enhancement features;
[0029] The shared feature enhancement unit uses the spatial enhancement features of infrared images and the spatial enhancement features of navigation radar images to obtain shared features, and enhances the shared features to obtain enhanced shared features;
[0030] The offset feature acquisition unit uses the enhanced shared features to acquire channel enhancement features, and links the channel enhancement features with infrared image features to obtain offset features;
[0031] The offset alignment unit uses offset features to align and compensate infrared image features, thereby obtaining aligned infrared image features.
[0032] Furthermore, the shared feature enhancement unit utilizes the spatial enhancement features of the infrared image and the spatial enhancement features of the navigation radar image to obtain shared features, and enhances the shared features to obtain enhanced shared features, specifically as follows:
[0033]
[0034]
[0035] in, It is an enhanced shared feature. It is a shared feature.
[0036] Furthermore, the offset feature acquisition unit utilizes the enhanced shared features to acquire channel enhancement features, and links the channel enhancement features with infrared image features to obtain offset features, specifically:
[0037]
[0038]
[0039] in, It is an offset feature. It is a channel enhancement feature. It is a multi-layer sensing network. It is an average pooling operation.
[0040] Furthermore, the offset alignment unit uses offset features to align and compensate for infrared image features, obtaining aligned infrared image features, specifically as follows:
[0041]
[0042] in, It's an alignment operation. Infrared image features pixel coordinates in It is an offset feature pixel coordinates in , Indicates at pixel point Pixel offset on Infrared image features The maximum value of the pixel's x-coordinate. Infrared image features The maximum value of the pixel's vertical coordinate.
[0043] Furthermore, the multi-stage feature fusion module includes: a feature fusion unit and a decoder unit;
[0044] The feature fusion unit is used to fuse aligned infrared image features and navigation radar image features to obtain fused features, specifically:
[0045]
[0046] in, It is a feature of fusion;
[0047] The decoder unit is used to aggregate the fused features to obtain aggregated features, and then use the aggregated features to obtain the iceberg target. Specifically:
[0048]
[0049]
[0050]
[0051] in, It is a batch normalization operation. , is a positive integer, , It is an aggregation feature. This is the decoder output.
[0052] Furthermore, the RIFANet network is trained using the training set, employing the following loss function:
[0053]
[0054]
[0055] in, It is the decoder's first One output, It's a real label. yes The median coordinate is pixels, yes The true label, express The pixel coordinates in the image.
[0056] The beneficial effects of this invention are as follows:
[0057] This invention proposes an infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction networks. The invention utilizes a dual-branch encoder to acquire complementary features from infrared and navigation radar images respectively. The detection modality branch acquires local and global features of small-sized targets at different levels. The reference modality branch avoids feature confusion and loss of detailed features due to the sparsity of radar features while ensuring the integrity of the learned features, thus improving the accuracy of iceberg target detection. The cross-modal feature alignment module designed in this invention calculates and aligns cross-modal offsets, promoting information exchange between modalities. Simultaneously, a multi-stage feature fusion module is designed to perform hierarchical feature fusion, reducing information redundancy caused by feature fusion at multiple resolutions, avoiding information differences between individual branches, and retaining effective features, further improving the accuracy of iceberg target detection. Attached Figure Description
[0058] Figure 1 This is a diagram of the iceberg detection network structure.
[0059] Figure 2 This is a structural diagram of the cross-modal feature alignment module;
[0060] Figure 3 A visualization comparing RIFANet's detection performance on the IRMIT dataset with existing methods. Detailed Implementation
[0061] Specific Implementation Method 1: This implementation method is based on an infrared-radar multimodal iceberg target detection method using heterogeneous feature alignment and information interaction networks.
[0062] The infrared image and the navigation radar image to be tested are acquired and then input into the trained RIFANet network to obtain the iceberg target detection result.
[0063] The trained RIFANet network is obtained through the following methods:
[0064] The infrared image of the iceberg target, the navigation radar image of the iceberg target, and the corresponding iceberg target detection results are combined into a training set. The RIFANet network is trained using the training set to obtain the trained RIFANet network.
[0065] like Figure 1 As shown, the RIFANet network includes: an encoder module, a cross-modal feature alignment module, and a multi-stage feature fusion module;
[0066] The encoder module uses infrared images and navigation radar images to obtain infrared image feature sets and navigation radar image feature sets;
[0067] The encoder module includes: a detection mode branch and a reference mode branch;
[0068] The detection modality branch utilizes infrared images to acquire infrared image feature sets. ;
[0069] The reference mode branch uses navigation radar images to obtain a set of navigation radar image features. ;
[0070] The cross-modal feature alignment module uses features from the infrared image feature set and the navigation radar image feature set to offset and align the infrared image features to obtain aligned infrared image features.
[0071] The multi-stage feature fusion module uses aligned infrared image features and infrared image feature sets to obtain iceberg target detection results.
[0072] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the detection modal branch includes: a first VGG unit, a second VGG unit, a third VGG unit, and a Transformer unit;
[0073] The input to the first VGG unit is an infrared image of an iceberg target, with dimensions of 512×512×3. The output of the first VGG unit is the infrared image features. The input to the second VGG unit is infrared image features. The output is infrared image features. The input to the third VGG unit is infrared image features. The output is infrared image features. The input to the Transformer unit is infrared image features. The output is infrared image features. ;
[0074] in, , , , .
[0075] The first VGG unit, the second VGG unit, and the third VGG unit acquire infrared image feature maps in the following ways. Specifically:
[0076]
[0077] in, It is an infrared image of an iceberg target. yes convolution, It is an activation function. It is a feature icon number; when hour, , It is an infrared image of an iceberg target;
[0078] The Transformer unit is a Transformer network, such as... Figure 1 As shown in (2) of the text;
[0079] The output of the detection modality branch is an infrared image feature set. ;
[0080] The reference mode branch includes: a first navigation radar image feature acquisition unit, a second navigation radar image feature acquisition unit, a third navigation radar image feature acquisition unit, and a fourth navigation radar image feature acquisition unit;
[0081] The first navigation radar image feature acquisition unit, the second navigation radar image feature acquisition unit, the third navigation radar image feature acquisition unit, and the fourth navigation radar image feature acquisition unit each include one A convolutional layer and a max-pooling layer;
[0082] The first navigation radar image feature acquisition unit takes a navigation radar image as input and outputs navigation radar image features. The second navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. The input to the third navigation radar image feature acquisition unit is the navigation radar image feature. The output is navigation radar image features. The fourth navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. ;
[0083] in, , , , .
[0084] The other steps and parameters are the same as in Specific Implementation Method 1.
[0085] In this embodiment, the detection modality branch and the reference modality branch are located in the shallow layers of the network, and their output feature maps contain rich shallow information, such as key structural features like edges and structures. As the network deepens, detailed features are gradually lost, and the deeper feature maps contain global features such as contrast features and large-scale structural features. The reference modality branch is used to process the PPI image of the navigation radar data. Since radar data is sparse, excessively deep network layers cannot extract features. Therefore, the radar data only goes through simple convolutional and pooling layers, and the size of the output feature map is similar to that output by the detection modality branch. Figure 1 One-to-one correspondence.
[0086] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: Figure 2 As shown, the cross-modal feature alignment module includes: a spatial enhancement feature acquisition unit, a shared feature enhancement unit, an offset feature acquisition unit, and an offset alignment unit;
[0087] The spatial enhancement feature acquisition unit is used to spatially enhance the infrared image features and navigation radar image features, and to acquire the spatial enhancement features of the infrared image and the navigation radar image, specifically:
[0088]
[0089]
[0090] in, It's about spatial awareness and operation. yes function, It's a link operation. It is a global max pooling operation in the spatial dimension. It is a spatial dimension average pooling operation. , It is a feature icon number. It's a multiplication of the channel dimension. It is the first infrared image Spatial enhancement features, It is the first navigation radar image Individual spatial enhancement features;
[0091] The shared feature enhancement unit utilizes the spatial enhancement features of infrared images and navigation radar images to obtain shared features, and enhances the shared features to obtain enhanced shared features, specifically as follows:
[0092]
[0093]
[0094] in, It is an enhanced shared feature. It is a shared feature;
[0095] The offset feature acquisition unit uses the enhanced shared features to acquire channel enhancement features, and links the channel enhancement features with infrared image features to obtain offset features, specifically:
[0096]
[0097]
[0098] in, It is an offset feature. It is a channel enhancement feature. It is a multi-layer sensing network. It is an average pooling operation;
[0099] The offset alignment unit uses offset features to align and compensate for infrared image features, obtaining aligned infrared image features, specifically:
[0100]
[0101] in, It's an alignment operation. Infrared image features pixel coordinates in It is an offset feature pixel coordinates in , Indicates at pixel point Pixel offset on Infrared image features The maximum value of the pixel's x-coordinate. Infrared image features The maximum value of the pixel's vertical coordinate.
[0102] The other steps and parameters are the same as in specific implementation method one or two.
[0103] In this embodiment, the output of the detection branch encoder will be used. and The cross-modal feature alignment module is input sequentially to process the inconsistencies between adjacent features step by step. The offset is calculated by extracting semantic features from two modalities at different resolutions. Then, feature alignment is performed based on the offset features. Finally, the aligned features are used to adaptively provide more accurate supplementary information for the detection modality. and Spatial attention operations are performed separately to obtain representations of the target and background in each modality, and each feature is enhanced. A shared feature enhancement unit multiplies the spatial enhancement features of two modalities pixel-by-pixel to obtain shared features. Then, the original features of the modality are added to the shared features to avoid inappropriate feature redundancy and enhance the shared features. After obtaining the enhanced shared features, a channel attention module further suppresses channel noise through differentiation. Different spatial context descriptions are calculated using average pooling and max pooling, and channel attention is obtained through a multilayer perceptron (MLP). Finally, the channel-enhanced features are fused again through a nonlinear MLP layer. The feature map of the detection modality branch is linked with the channel-enhanced features to obtain the calculated modality shift features. Based on the shift features, an alignment function is used to adaptively align and compensate the features. The alignment function uses the obtained shift features to perform bilinear interpolation on the feature map of the detection modality, further transforming the features, and finally obtaining the aligned infrared image features.
[0104] Specific Implementation Method Four: This implementation method differs from one of Specific Implementation Methods One to Three in that the multi-stage feature fusion module includes: a feature fusion unit and a decoder unit;
[0105] The feature fusion unit is used to fuse aligned infrared image features and navigation radar image features to obtain fused features, specifically:
[0106]
[0107] in, It is a feature of fusion;
[0108] The decoder unit is used to aggregate the fused features to obtain aggregated features, and then use the aggregated features to obtain the iceberg target. Specifically:
[0109]
[0110]
[0111]
[0112] in, It is a batch normalization operation. , is a positive integer, , It is an aggregation feature. It is the decoder output;
[0113] The other steps and parameters are the same as in any of the specific implementation methods one to three.
[0114] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One through Four in that: the RIFANet network is trained using the training set, and the following loss function is employed:
[0115]
[0116]
[0117] in, It is the decoder's first One output, It's a real label. yes The median coordinate is pixels, yes The true label, express pixel coordinates in .
[0118] The Soft-IoU loss function in this step can effectively handle class imbalance and small target problems, helping the network to segment small targets in infrared scenes and perform fine segmentation of targets at the pixel level.
[0119] The other steps and parameters are the same as in any of the specific implementation methods one to four.
[0120] Example: To verify the beneficial effects of the present invention, this example compares the present invention with existing methods, specifically as follows:
[0121] As shown in Tables 1 and 2, this invention demonstrates significant superiority over other methods in terms of performance metrics. In the river scene of the IRMIT infrared multimodal iceberg target dataset, RIFANet achieved optimal results in IoU, nIoU, Pd, and F1 scores, far exceeding other methods. Furthermore, in the comparative experiment on the IRMIT sea surface scene, the evaluation metrics IoU, nIoU, Pd, and F1 also ranked first among all metrics. Qualitative analysis reveals that RIFANet exhibits outstanding advantages in both infrared iceberg target detection and infrared small target detection. Figure 3 The visualization comparison shows that the present invention (RIFANet) emphasizes and aligns the feature advantages of the two modes by referencing the supplementary information of the mode, effectively avoiding the interference of poor visual conditions and clutter. For weak iceberg targets in complex scenes, our method shows stronger robustness.
[0122] Table 1
[0123]
[0124] Table 2
[0125]
Claims
1. An infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network, characterized in that... The specific process of the method is as follows: Acquire the infrared image and the navigation radar image to be tested, and input the infrared image and the navigation radar image to be tested into the trained RIFANet network to obtain the iceberg target detection result; The RIFANet network includes: an encoder module, a cross-modal feature alignment module, and a multi-stage feature fusion module; The encoder module uses infrared images and navigation radar images to obtain infrared image feature sets and navigation radar image feature sets; The cross-modal feature alignment module uses features from the infrared image feature set and the navigation radar image feature set to offset and align the infrared image features to obtain aligned infrared image features. The multi-stage feature fusion module uses aligned infrared image features and infrared image feature sets to obtain iceberg target detection results.
2. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 1, characterized in that: The trained RIFANet network is obtained through the following method: The infrared image of the iceberg target, the navigation radar image of the iceberg target, and the corresponding iceberg target detection results are combined into a training set. The RIFANet network is trained using the training set to obtain the trained RIFANet network.
3. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 2, characterized in that: The encoder module includes: a detection mode branch and a reference mode branch; The detection modality branch utilizes infrared images to acquire infrared image feature sets. ; The detection modality branch includes: a first VGG unit, a second VGG unit, a third VGG unit, and a Transformer unit; The input to the first VGG unit is an infrared image of the iceberg target, and the output of the first VGG unit is the infrared image features. The input to the second VGG unit is infrared image features. The output is infrared image features. The input to the third VGG unit is infrared image features. The output is infrared image features. The input to the Transformer unit is infrared image features. The output is infrared image features. ; The reference mode branch uses navigation radar images to obtain a set of navigation radar image features. ; The reference mode branch includes: a first navigation radar image feature acquisition unit, a second navigation radar image feature acquisition unit, a third navigation radar image feature acquisition unit, and a fourth navigation radar image feature acquisition unit; The first navigation radar image feature acquisition unit, the second navigation radar image feature acquisition unit, the third navigation radar image feature acquisition unit, and the fourth navigation radar image feature acquisition unit each include one A convolutional layer and a max-pooling layer; The first navigation radar image feature acquisition unit takes a navigation radar image as input and outputs navigation radar image features. The second navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. The third navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. The fourth navigation radar image feature acquisition unit receives navigation radar image features as input. The output is navigation radar image features. .
4. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 3, characterized in that: The infrared image features , , Obtained through the following methods: in, It is an infrared image of an iceberg target. yes convolution, It is an activation function. It is a feature icon number; when hour, , This is an infrared image of an iceberg target.
5. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 4, characterized in that: The cross-modal feature alignment module includes: a spatial enhancement feature acquisition unit, a shared feature enhancement unit, an offset feature acquisition unit, and an offset alignment unit; The spatial enhancement feature acquisition unit is used to spatially enhance the infrared image features and navigation radar image features, and to acquire the spatial enhancement features of the infrared image and the navigation radar image, specifically: in, It's about spatial awareness and operation. yes function, It's a link operation. It is a global max pooling operation in the spatial dimension. It is a spatial dimension average pooling operation. , It is a feature icon number. It's a multiplication of the channel dimension. It is the first infrared image A spatial enhancement feature, It is the first navigation radar image Individual spatial enhancement features; The shared feature enhancement unit uses the spatial enhancement features of infrared images and the spatial enhancement features of navigation radar images to obtain shared features, and enhances the shared features to obtain enhanced shared features; The offset feature acquisition unit uses the enhanced shared features to acquire channel enhancement features, and links the channel enhancement features with infrared image features to obtain offset features; The offset alignment unit uses offset features to align and compensate infrared image features, thereby obtaining aligned infrared image features.
6. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 5, characterized in that: The shared feature enhancement unit utilizes the spatial enhancement features of infrared images and navigation radar images to obtain shared features, and enhances the shared features to obtain enhanced shared features, specifically as follows: in, It is an enhanced shared feature. It is a shared feature.
7. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 6, characterized in that: The offset feature acquisition unit uses the enhanced shared features to acquire channel enhancement features, and links the channel enhancement features with infrared image features to obtain offset features, specifically: in, It is an offset feature. It is a channel enhancement feature. It is a multi-layer sensing network. It is an average pooling operation.
8. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 7, characterized in that: The offset alignment unit uses offset features to align and compensate for infrared image features, obtaining aligned infrared image features, specifically: in, It's an alignment operation. Infrared image features pixel coordinates in It is an offset feature pixel coordinates in , Indicates at pixel point Pixel offset on Infrared image features The maximum value of the pixel's x-coordinate. Infrared image features The maximum value of the pixel's vertical coordinate.
9. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 8, characterized in that: The multi-stage feature fusion module includes: a feature fusion unit and a decoder unit; The feature fusion unit is used to fuse aligned infrared image features and navigation radar image features to obtain fused features, specifically: in, It is a feature of fusion; The decoder unit is used to aggregate the fused features to obtain aggregated features, and to use the aggregated features to obtain the iceberg target. Specifically: in, It is a batch normalization operation. , is a positive integer, , It is an aggregation feature. This is the decoder output.
10. The infrared-radar multimodal iceberg target detection method based on heterogeneous feature alignment and information interaction network according to claim 9, characterized in that: The RIFANet network is trained using the training set, employing the following loss function: in, It is the decoder's first One output, It's a real label. yes The median coordinate is pixels, yes The true label, express The pixel coordinates in the image.