Infrared ship identification method and system based on wavelet transform deformable convolution and efficient additional attention

Abstract

The invention provides an infrared ship identification method and system based on wavelet transform deformable convolution and efficient additional attention, and belongs to the field of deep learning. Creating a model based on a classical target detection algorithm RT-DETR principle, reproducing initialization, and constructing and processing an infrared marine ship data set; an RT-DETR target detection algorithm is improved into a wavelet transform deformable convolution and efficient additional attention network model architecture, an infrared marine ship data set after data processing is used in a model framework, related parameters are debugged to perform iterative training on the wavelet transform deformable convolution and efficient additional attention network model, and an expected test effect is obtained; in order to expand the application range of the algorithm, the model is packaged, stored and deployed to an artificial intelligence development board. According to the method, the problems of low target contrast, large complex sea clutter influence, multi-scale target shielding and the like in the infrared image can be solved, and the ship target detection precision and robustness in the complex marine environment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an infrared ship recognition method and system based on wavelet transform deformable convolution and efficient additional attention. Background Art

[0002] In the modern marine environment, maritime traffic is busy and complex, posing severe challenges to the country's marine resource management and border security. Strengthening maritime monitoring and target detection has become a top priority. At present, maritime monitoring mostly relies on traditional optical equipment, but the performance of traditional equipment is significantly reduced at night or in bad weather conditions. However, with the rapid development of deep learning in the field of computer vision, neural network-based target detection algorithms have shown strong application potential in the fields of target recognition in complex environments, maritime monitoring, and infrared imaging, providing a new technical path to solve the above problems.

[0003] Therefore, it is of great significance to realize automatic detection of ship targets in complex marine environments. Infrared ship target detection is part of the target detection field, which aims to accurately locate and classify ships at sea in complex backgrounds, thereby improving the effect of maritime safety monitoring. At present, the main target detection algorithms rely on deep learning technology, which designs convolutional neural networks and uses large-scale labeled image data for training to cope with target detection and recognition tasks in complex scenes. With the continuous development of deep learning, two-stage target detection algorithms based on candidate boxes, such as the RCNN series, have gradually been widely used. This type of algorithm first generates candidate regions, then classifies and corrects them by position regression, and finally obtains the detection results. However, this algorithm shows the problem of insufficient computational efficiency in scenes with high real-time requirements, because feature extraction and classification regression are processed separately. To this end, researchers have proposed regression-based one-stage target detection algorithms, such as the YOLO series, which achieve faster detection speed and higher efficiency by directly determining the target position in the entire image. With the development of the Transformer architecture in deep learning, researchers have begun to explore its application in target detection, providing new solutions for small target detection in complex scenes.

[0004] DETR (Detection Transformer) is a new end-to-end object detection algorithm. It breaks through the traditional candidate box-based object detection mode, adopts the Transformer architecture, and cancels the steps such as the region proposal network (RPN) and selective search. The operating principle of DETR is to directly map the input image to a set of fixed number of target queries, and use the Transformer to perform global feature modeling and relationship capture on these queries, so as to achieve object detection and classification. In DETR, the entire image is first extracted by a convolutional neural network (such as ResNet), and then the Transformer performs global modeling on these features through position encoding and self-attention mechanism to capture the relationship between objects. Finally, after processing by the multi-head attention mechanism and the fully connected layer, the category and position of each object are output. A significant advantage of DETR is that it simplifies the object detection process, integrates all steps into an end-to-end framework, and no longer relies on the generation and screening of candidate boxes. Compared with the traditional two-stage algorithm, DETR has shown excellent performance in both detection speed and accuracy. However, DETR has certain limitations when dealing with small and dense targets, mainly due to its low attention to details in the global modeling process.

[0005] RT-DETR has made two important improvements to solve the convergence problem of small target recognition: Efficient Hybrid Encoder and IoU-aware Query Module. The Efficient Hybrid Encoder significantly improves the model's ability to extract small target features at different scales and complex scenes by integrating multiple feature encoding methods. At the same time, this encoder also optimizes computational efficiency, allowing RT-DETR to reduce computational costs while maintaining high accuracy. The IoU-aware Query Module further enhances the ability to distinguish between targets and backgrounds. By dynamically adjusting the position and size of query points, the module enables the model to better capture the overlap between targets and backgrounds, thereby improving the detection accuracy of small targets and targets with complex shapes. These two improvements enable RT-DETR to perform well in complex scenes and multi-scale target detection, while maintaining the global modeling capabilities and end-to-end simplicity of the DETR model.

[0006] Although RT-DETR target detection has a relatively high accuracy in traditional natural image recognition, there are still many problems in recognizing unnatural images such as infrared marine ship images. Due to the internal structure limitations of the classic RT-DETR's backbone network, encoder, decoder, etc., some problems will arise when recognizing unconventional images.

[0007] For example, when using the RT-DETR target detection algorithm to process infrared ship images in complex ocean environments, the following problems may be encountered:

[0008] 1. Feature extraction is not precise enough and has high spatial complexity: Although RT-DETR introduces deformable convolution to improve the ability to capture local features, the accuracy of feature extraction is still insufficient when dealing with multi-scale, small-target ship detection. Especially in complex ocean backgrounds, due to the low contrast and high noise of infrared images, simple convolution operations may not be able to effectively extract subtle target features. In addition, the feature extraction process of RT-DETR involves a large number of sampling point calculations, which increases the spatial complexity of the calculation to a certain extent, resulting in a decrease in computational efficiency.

[0009] 2. Limitations of multi-scale deformed target processing: Although RT-DETR has improved the detection capability of targets of different scales by introducing flexible sampling points, the recognition accuracy may still decrease when dealing with multi-scale deformed targets in infrared ship images. Especially in the face of ship targets with small pixels and multiple deformations in the offshore scene, a single static feature receptive field may not be able to fully capture the detailed information of the target, resulting in missed detection or false detection of the target.

[0010] 3. Insufficient accuracy in small target ship detection: Although RT-DETR alleviates the problem of slow convergence of small target detection by introducing single-scale feature interaction and cross-scale fusion modules, the feature map it uses is too deep to map the ship targets that are extremely small in scale. There is no design module within the algorithm that specifically focuses on the key features of small targets, which makes it impossible to accurately identify small target ships in the distant sea environment.

[0011] At present, there are the following problems in infrared ship target detection:

[0012] (1) There are a limited number of deep learning algorithms specifically designed for infrared ship target detection.

[0013] (2) The background of infrared ship target detection is complex, and small targets and deformation problems are difficult to handle.

[0014] (3) There are almost no deployment solutions for the above algorithms. Summary of the invention

[0015] In view of this, the present invention discloses an infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention. By combining an improved RT-DETR algorithm with wavelet transform deformable convolution, efficient additional attention and other technologies, high-precision detection of ship targets in complex infrared ocean environments is achieved, and the model can be deployed on an artificial intelligence development board.

[0016] On the one hand, the present invention provides an infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention, comprising:

[0017] Obtain an infrared ship image dataset, perform data enhancement on the image dataset, annotate and classify the locations and categories of the ships in the image dataset, and divide the image dataset into a training set and a test set;

[0018] An infrared ship recognition model based on an improved RT-DETR target detection network is constructed. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. The RT-DETR target detection model is initialized, and the C2-C4 ordinary convolution blocks in the original backbone network Resnet50 are replaced by the wavelet transform WTConv module. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution. The deformable convolution DCNv4-CVPR2024 module is fused to the original feature extraction C5 layer, and the encoder small target focusing module is added. The original encoder AIFI stage is replaced by the efficient additional attention EfficientAdditiveAttention module, and the small target detection module P2 is introduced backward to complete the network reconstruction.

[0019] Using the training set to train the infrared ship recognition model;

[0020] The trained infrared ship recognition model is used to detect and identify sub-classified ship targets in infrared images.

[0021] Furthermore, after using the training set to train the infrared ship recognition model, it also includes: using the test set to test the infrared ship recognition model, calculating the generalization ability of the infrared ship recognition model, and selecting and saving the trained infrared ship recognition model according to the generalization ability.

[0022] Furthermore, after obtaining the trained infrared ship recognition model, the method further includes: packaging the trained infrared ship recognition model, configuring the operating environment of the model on the artificial intelligence development board and testing it;

[0023] The artificial intelligence development board is used to detect and identify sub-classified ship targets in infrared images.

[0024] Furthermore, the construction of the WTConv module includes:

[0025] After wavelet decomposition, the input feature map is transformed into a low-frequency part and a high-frequency part;

[0026] After the convolution operation, the low-frequency and high-frequency components are convolved separately;

[0027] The feature map after the convolution operation is reconstructed by inverse wavelet transformation to reconstruct the final output feature map.

[0028] Furthermore, the deformable convolution realizes feature extraction of dynamic adaptive deformable targets by introducing offset learning and dynamic sampling;

[0029] The deformable convolution DCNv4-CVPR2024 module is fused into the original feature extraction C5 layer, including:

[0030] Design a DCNv4 offset learning network that predicts the offset of each convolution position through an additional convolutional layer, and the convolution kernel dynamically adjusts its sampling position according to the offset;

[0031] Apply the DCNv4 offset learning network to the feature map X of layer C5.

[0032] Furthermore, the EfficientAdditiveAttention module optimizes the Attention operation into an additive form by reducing the computational complexity.

[0033] Furthermore, the P2 module is specifically used to process detail information related to small targets in the feature map, and ensures that small targets will not be missed or misdetected in the final detection process by performing multi-scale fusion and amplification of features in the small target area.

[0034] In yet another aspect, the present invention further discloses an infrared ship recognition system based on wavelet transform deformable convolution and efficient additional attention, comprising: a data set creation module, a model construction module, a model training module and a target recognition module;

[0035] The dataset creation module is used to obtain an image dataset of infrared ships, perform data enhancement on the image dataset, annotate and divide the positions and categories of ships in the image dataset, and divide the image dataset into a training set and a test set;

[0036] The model building module is used to build an infrared ship recognition model based on the improved RT-DETR target detection network. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. The RT-DETR target detection model is initialized, and the C2-C4 ordinary convolution blocks in the original backbone network Resnet50 are replaced by the wavelet transform WTConv module. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution; the deformable convolution DCNv4-CVPR2024 module is fused to the original feature extraction C5 layer, and the encoder small target focusing module is added. The original encoder AIFI stage is replaced by the efficient additional attention EfficientAdditiveAttention module, and the small target detection module P2 is introduced backward to complete the network reconstruction;

[0037] The model training module is used to train the infrared ship recognition model using the training set obtained by the data set creation module, and to test the infrared ship recognition model using the test set, to calculate the generalization ability of the infrared ship recognition model, and to select and save the trained infrared ship recognition model according to the generalization ability;

[0038] The target recognition module is used to detect and identify ship sub-classification targets in infrared images using the trained infrared ship recognition model.

[0039] Furthermore, it also includes:

[0040] The deployment module is used to encapsulate the saved infrared ship recognition model, configure the operating environment of the model on the artificial intelligence development board and perform tests to detect and identify infrared ships.

[0041] On the other hand, the present invention also discloses an artificial intelligence development board deployed with an infrared ship recognition model, wherein the infrared ship recognition model is a trained infrared ship recognition model obtained in the above-mentioned infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention.

[0042] The present invention deploys the wavelet transform deformable convolution and the infrared ship recognition method with efficient additional attention to the artificial intelligence development board, realizes the deployment of the model, and expands the application scope of the model.

[0043] The present invention aims to propose an infrared ship recognition method and deployment based on wavelet transform deformable convolution and efficient additional attention to enhance the detection capability in maritime security monitoring and military applications. By improving the algorithm and the actual deployment platform, the present invention solves the shortcomings of traditional methods in complex environments, provides stronger detection accuracy and real-time performance, and has wide application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 It is a flow chart of an infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention in an embodiment of the present invention;

[0046] Figure 2 It is a structural diagram of an infrared ship recognition system based on wavelet transform deformable convolution and efficient additional attention in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the deployment test results of the artificial intelligence development board in an embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of the structure of the RT-DETR target detection algorithm in an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of the improved structure of the RT-DETR target detection algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0052] like Figure 1 As shown, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention in an embodiment of the present invention comprises the following steps:

[0053] S1. Acquire an infrared ship image dataset, mark and classify the location and category information of the ship, and perform data enhancement and image preprocessing on the image dataset;

[0054] Among them, data enhancement can use the traditional RT-DETR source code to perform operations such as image translation, rotation, and cropping on the data set to achieve the purpose of image enhancement and expansion, and divide the image data set into a training set and a test set;

[0055] S2. Construct an infrared ship recognition model based on an improved RT-DETR target detection network. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. Initialize the RT-DETR target detection model, and use the wavelet transform WTConv module to replace the C2-C4 ordinary convolution blocks in the original backbone network Resnet50. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution; fuse the deformable convolution DCNv4-CVPR2024 module to the original feature extraction C5 layer, add the encoder small target focusing module, use the efficient additional attention EfficientAdditiveAttention module to replace the original encoder AIFI stage, and introduce the small target detection module P2 backward to complete the network reconstruction;

[0056] like Figure 4 , Figure 5 As shown in the figure, wavelet transform is a signal processing technology that extracts the features of the signal at different scales by decomposing the signal into different frequency bands. It has the ability of multi-resolution analysis and is very suitable for processing the features in high-noise ocean scenes. In order to use the wavelet transform convolution module (WTConv) in the C2-C4 layer of Resnet50, it is necessary to first design the WTConv module, which replaces the standard convolution operation with the operation combining wavelet transform and convolution. Then, after the convolution operation, the low-frequency and high-frequency components are convolved respectively, and finally the feature map Y after the convolution operation is reconstructed by inverse wavelet. l and Y h An inverse wavelet transform (IWT) is performed to reconstruct the final output feature map Y, and the obtained WTConv module replaces the C2-C4 ordinary convolution blocks in the original backbone network Resnet50.

[0057] Deformable convolution is an enhanced convolution operation that aims to dynamically adjust the position of the convolution kernel so that it can adapt to the shape and size of the target more flexibly. In standard convolution, the position of the convolution kernel is fixed, that is, for each position on the input feature map, the convolution kernel will slide and calculate according to predefined rules. However, for targets with complex shapes or large scale changes, fixed convolution kernels may not be able to effectively capture their features. Deformable convolution introduces offset learning and dynamic sampling to achieve dynamic adaptive deformation targets and extract features. In order to integrate the deformable convolution DCNv4-CVPR2024 module into the original feature extraction C5 layer, Deformable convolution is an enhanced convolution operation that aims to dynamically adjust the position of the convolution kernel so that it can adapt to the shape and size of the target more flexibly. In standard convolution, the position of the convolution kernel is fixed, that is, for each position on the input feature map, the convolution kernel will slide and calculate according to predefined rules. However, for targets with complex shapes or large scale changes, fixed convolution kernels may not be able to effectively capture their features. Deformable convolution achieves feature extraction of dynamically adaptive deformable targets by introducing offset learning and dynamic sampling. When the deformable convolution DCNv4-CVPR2024 module is fused into the original feature extraction C5 layer, a DCNv4 offset learning network is first designed, which predicts the offset Δp of each convolution position through an additional convolution layer. k The convolution kernel dynamically adjusts its sampling position according to these offsets, and the convolution operation becomes:

[0058]

[0059] Among them, y(p) is the value of the output feature map at position p, W k is the weight of the convolution kernel, x(p+p k +Δp k ) is obtained by the offset Δp k The sampled values ​​of the resized input feature map.

[0060] Then the DCNv4 offset learning network is applied to the feature map X of the C5 layer, and the formula is expressed as:

[0061]

[0062] Among them, Y C5 is the output feature map of layer C5, W C5 is the convolution kernel weight of the C5 layer. By fusing the deformable convolution module, the C5 layer can dynamically adapt to the shape and scale changes of the target, effectively improving the detection ability of complex targets. Finally, the fused feature map is combined with the feature maps of other layers to form a complete feature extraction network.

[0063] The EfficientAdditiveAttention (EAA) module aims to improve the model's ability to detect small targets by effectively focusing on key areas in the input features. In the encoder stage, by using the EAA module, the feature extraction capability of small targets can be enhanced and feature loss caused by background noise or scale changes can be reduced. Specifically, EfficientAdditiveAttention optimizes the Attention operation into an additive form by reducing computational complexity. The formula is:

[0064] EAA(Q,K,V)=V+α·σ(Q+K);

[0065] Here, α is a learnable scalar weight and σ is a nonlinear activation function.

[0066] In the original encoder AIFI stage, it is replaced with the EAA module to enhance the model's ability to focus on important features. The EAA module can not only effectively focus on key features in the global scope, but also dynamically adjust the focus on small targets, thereby improving detection accuracy. At the same time, in order to further improve the detection performance of small targets, the small target detection module P2 is introduced after the EAA module. The P2 module is specifically used to process the detail information related to small targets in the feature map. By multi-scale fusion and amplification of the features of the small target area, it ensures that small targets will not be missed or misdetected in the final detection process.

[0067] like Figure 2 As shown, an embodiment of the present invention also provides an infrared ship recognition system based on wavelet transform deformable convolution and efficient additional attention, including: a data set creation module, a model construction module, a model training module and a target recognition module;

[0068] The dataset creation module is used to obtain an image dataset of infrared ships, perform data enhancement on the image dataset, annotate and divide the positions and categories of ships in the image dataset, and divide the image dataset into a training set and a test set;

[0069] The model building module is used to build an infrared ship recognition model based on the improved RT-DETR target detection network. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. The RT-DETR target detection model is initialized, and the C2-C4 ordinary convolution blocks in the original backbone network Resnet50 are replaced by the wavelet transform WTConv module. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution; the deformable convolution DCNv4-CVPR2024 module is fused to the original feature extraction C5 layer, and the encoder small target focusing module is added. The original encoder AIFI stage is replaced by the efficient additional attention EfficientAdditiveAttention module, and the small target detection module P2 is introduced backward to complete the network reconstruction;

[0070] The model training module is used to train the infrared ship recognition model using the training set obtained by the data set creation module, and to test the infrared ship recognition model using the test set, to calculate the generalization ability of the infrared ship recognition model, and to select and save the trained infrared ship recognition model according to the generalization ability;

[0071] The target recognition module is used to detect and identify the ship sub-classification targets in the infrared image using the trained infrared ship recognition model;

[0072] The deployment module is used to encapsulate the saved RT-DETR target detection model, configure the operating environment of the model on the artificial intelligence development board and perform tests to detect and identify infrared ships.

[0073] The experimental verification was carried out in a single GTX 3090 hardware environment and Python software environment. The simulation experiment of the present invention showed the image test results based on the deployment of the artificial intelligence development board. Figure 3 It can be seen that the model can realize accurate infrared image ship classification and recognition on the artificial intelligence development board, among which fishing boat 0.92 is an identified fishing boat and sailboat 0.95 is an identified sailboat.

[0074] The present invention provides an infrared ship recognition method and deployment based on wavelet transform deformable convolution and efficient additional attention. It is based on the principle of RT-DETR target detection algorithm, and optimizes and improves the backbone network in its network structure. The wavelet transform WTConv module is used to replace the C2-C4 ordinary convolution blocks in the original backbone network Resnet50, thereby avoiding over-parameterization when extracting multi-resolution ship features in complex and noisy ocean infrared scenes, and obtaining a better global optimization receptive field.

[0075] At the same time, the neck network in the network structure based on the RT-DETR target detection model is optimized and improved. By integrating the deformable convolution DCNv4-CVPR2024 module into the C5 layer, the algorithm can more flexibly adjust the receptive field and adaptively capture the ship features in complex scenes, especially in the detection of multi-scale and irregular targets, showing higher accuracy.

[0076] The encoder in the network structure based on the RT-DETR target detection model is optimized and improved, and the EfficientAdditiveAttention module is used to replace the original encoder AIFI stage. At the same time, the small target detection module P2 is introduced backward, so as to obtain a more effective feature focusing mechanism to optimize the extraction of key features of small targets and thus increase accuracy.

[0077] The present invention deploys the wavelet transform deformable convolution and the infrared ship recognition method with efficient additional attention to the artificial intelligence development board, realizes the deployment of the model, and expands the application scope of the model.

[0078] The present invention deploys the hyperparameter-based residual convolution and clustering fusion algorithm to the artificial intelligence development board, realizes the deployment of the model, and expands the application scope of the model.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention, characterized in that: include: Obtain an infrared ship image dataset, perform data enhancement on the image dataset, annotate and classify the locations and categories of the ships in the image dataset, and divide the image dataset into a training set and a test set; An infrared ship recognition model based on an improved RT-DETR target detection network is constructed. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. The RT-DETR target detection model is initialized, and the C2-C4 ordinary convolution blocks in the original backbone network Resnet50 are replaced by the wavelet transform WTConv module. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution. The deformable convolution DCNv4-CVPR2024 module is fused to the original feature extraction C5 layer, and the encoder small target focusing module is added. The original encoder AIFI stage is replaced by the efficient additional attention EfficientAdditiveAttention module, and the small target detection module P2 is introduced backward to complete the network reconstruction. Using the training set to train the infrared ship recognition model; The trained infrared ship recognition model is used to detect and identify sub-classified ship targets in infrared images.

2. According to claim 1, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: After using the training set to train the infrared ship recognition model, the method also includes: using the test set to test the infrared ship recognition model, calculating the generalization ability of the infrared ship recognition model, and selecting and saving the trained infrared ship recognition model according to the generalization ability.

3. According to claim 2, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: After obtaining the trained infrared ship recognition model, the method further includes: packaging the trained infrared ship recognition model, configuring the operating environment of the model on the artificial intelligence development board, and testing the model; The artificial intelligence development board is used to detect and identify sub-classified ship targets in infrared images.

4. According to claim 1, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: The construction of the WTConv module includes: After wavelet decomposition, the input feature map is transformed into a low-frequency part and a high-frequency part; After the convolution operation, the low-frequency and high-frequency components are convolved separately; The feature map after the convolution operation is reconstructed by inverse wavelet transformation to reconstruct the final output feature map.

5. According to claim 1, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: The deformable convolution realizes feature extraction of dynamic adaptive deformation targets by introducing offset learning and dynamic sampling; The deformable convolution DCNv4-CVPR2024 module is fused into the original feature extraction C5 layer, including: Design a DCNv4 offset learning network that predicts the offset of each convolution position through an additional convolutional layer, and the convolution kernel dynamically adjusts its sampling position according to the offset; Apply the DCNv4 offset learning network to the feature map X of layer C5.

6. According to claim 1, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: The EfficientAdditiveAttention module optimizes the Attention operation into an additive form by reducing the computational complexity.

7. According to claim 1, the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention is characterized in that: The P2 module is specifically used to process the detail information related to the small target in the feature map, and ensures that the small target will not be missed or misdetected in the final detection process by fusing and amplifying the features of the small target area at multiple scales.

8. An infrared ship recognition system based on wavelet transform deformable convolution and efficient additional attention, characterized in that: include: Dataset creation module, model building module, model training module and target recognition module; The dataset creation module is used to obtain an image dataset of infrared ships, perform data enhancement on the image dataset, annotate and divide the positions and categories of ships in the image dataset, and divide the image dataset into a training set and a test set; The model building module is used to build an infrared ship recognition model based on the improved RT-DETR target detection network. The improved RT-DETR target detection network includes a backbone network, an encoder, and a decoder. The RT-DETR target detection model is initialized, and the wavelet transform WTConv module is used to replace the C2-C4 ordinary convolution blocks in the original backbone network Resnet50. The WTConv module replaces the standard convolution operation with an operation combining wavelet transform and convolution; the deformable convolution DCNv4-CVPR2024 module is fused to the original feature extraction C5 layer, and the encoder small target focusing module is added. The efficient additional attention EfficientAdditiveAttention module is used to replace the original encoder AIFI stage, and the small target detection module P2 is introduced backward to complete the network reconstruction; The model training module is used to train the infrared ship recognition model using the training set obtained by the data set creation module, and to test the infrared ship recognition model using the test set, to calculate the generalization ability of the infrared ship recognition model, and to select and save the trained infrared ship recognition model according to the generalization ability; The target recognition module is used to detect and identify ship sub-classification targets in infrared images using the trained infrared ship recognition model.

9. The infrared ship recognition system based on wavelet transform deformable convolution and efficient additional attention according to claim 8 is characterized in that: Also includes: The deployment module is used to encapsulate the saved infrared ship recognition model, configure the operating environment of the model on the artificial intelligence development board and perform tests to detect and identify infrared ships.

10. An artificial intelligence development board deployed with an infrared ship recognition model, wherein the infrared ship recognition model is a trained infrared ship recognition model obtained from the infrared ship recognition method based on wavelet transform deformable convolution and efficient additional attention as described in any one of claims 1 to 8.

Images