Lightweight real-time infrared target detection method and system for unmanned ship

Abstract

The invention relates to the technical field of infrared target detection, in particular to a light-weight real-time infrared target detection method and system for an unmanned ship, and the method comprises the steps: collecting infrared data through an infrared thermal imager carried by the unmanned ship, and carrying out the preprocessing, marking and dividing. And a lightweight GLAF-MobileNetV4 network is adopted as a backbone of the detection method, so that the real-time performance is improved, and model parameters are greatly reduced. Wherein the lightweight global and local adaptive fusion attention can ingeniously integrate global and local information, and local information fusion is adjusted and optimized through dynamic weight. Meanwhile, the designed dynamic adaptive multi-scale feature fusion encoder adaptively adjusts the fusion proportion according to the importance of the multi-scale features by using an attention guidance mechanism. The method is simple and convenient to deploy, improves the target detection precision and real-time performance of the unmanned ship under the low-light condition, and is remarkable in detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of infrared target detection technology, and in particular to a lightweight real-time infrared target detection method and system for unmanned boats. Background Art

[0002] With the rapid development of artificial intelligence, autonomous navigation and communication technology, the application prospects of unmanned boats are becoming more and more broad. The research of unmanned boats is of great significance to enhancing the country's maritime strength and strengthening national defense security. With a high degree of autonomy and intelligence, unmanned boats can independently complete a series of complex tasks. Especially in low-light conditions, in order to ensure that unmanned boats can continue to operate autonomously, it is particularly important to use infrared thermal imagers to capture sea scene information. Therefore, infrared target detection technology has become a key core technology for unmanned boats to efficiently perform tasks in low-light environments.

[0003] In complex and ever-changing sea scenes, infrared target detection technology needs to be lightweight and real-time to adapt to platform environments with limited computing power and to quickly and accurately detect and identify sea surface targets. However, the feature information of infrared sea surface targets is relatively limited and lacks rich texture details, which undoubtedly increases the difficulty of detection. In addition, the vast sea surface often leads to targets of different sizes appearing in infrared images, and the significant size differences between targets further increase the risk of false detection and missed detection.

[0004] In order to improve the real-time performance of infrared target detection technology, the common practice is to use a lightweight backbone network combined with the YOLO algorithm framework. However, the YOLO algorithm relies on the non-maximum suppression method to eliminate redundant prediction boxes. This post-processing step often requires manual parameter adjustment, which not only affects the stability of detection accuracy, but may also slow down the detection speed. On the other hand, although the lightweight backbone network can effectively reduce the computational burden, it is often accompanied by a decrease in detection accuracy. In order to make up for the shortcomings, an attention mechanism is usually introduced to enhance the salient features of infrared targets while suppressing background noise, thereby improving the recognition of target information. However, existing attention mechanisms, such as Squeeze-Excitation (SE) channel attention and Convolutional Block Attention Module (CBAM), still have certain limitations in detection accuracy and stability when applied to infrared sea surface target detection. Summary of the invention

[0005] In view of this, the purpose of the present invention is to propose a lightweight real-time infrared target detection method and system for unmanned boats, so as to solve the shortcomings of existing infrared sea surface target detection methods in real-time and detection effect.

[0006] Based on the above purpose, the present invention provides a lightweight real-time infrared target detection method for unmanned boats, comprising the following steps:

[0007] S1. Collect infrared video data of the unmanned boat surface environment, perform preprocessing, construct an infrared image target data set, annotate and classify each target in the data set, and divide the annotated data set into a training set, a validation set, and a test set according to a certain ratio;

[0008] S2. Train the lightweight real-time infrared target detection model through the data set to obtain a trained model, wherein the detection process of the lightweight real-time infrared target detection model includes:

[0009] Using the lightweight backbone network GLAF-MobileNetV4 based on the Global Local Adaptive Fusion (GLAF) attention mechanism, deep feature extraction is performed on infrared images to capture multi-scale and multi-level feature information in the image.

[0010] The multi-scale and multi-level feature information is fed into a dynamic adaptive multi-scale feature fusion encoder, and the minimum uncertainty query selection method is used to select key features from the encoder output as the initial object query;

[0011] The Transformer-based decoder refines these initial object queries to obtain target boxes and corresponding category confidences.

[0012] S3. Detect the target to be detected through the trained lightweight real-time infrared target detection model to obtain the detection result.

[0013] Preferably, in step S1, the scenes of the collected infrared video data of the unmanned boat water surface environment include a variety of water environments, the time period for data collection covers a plurality of time periods, the weather conditions include different weather conditions, and the motion state of the unmanned boat includes a variety of motion states.

[0014] Preferably, the lightweight backbone network GLAF-MobileNetV4 includes five stages, the first three stages include convolutional layers, and the GLAF attention mechanism is added at the end of the second and third stages, the fourth and fifth stages include ExtraDepthwise, Inverted Bottleneck and GLAF-ConvNext modules, among which the GLAF-ConvNext module embeds the GLAF mechanism, and the Extra Depthwise, Inverted Bottleneck and GLAF-ConvNext modules all adopt the Universal Inverted Bottleneck structure.

[0015] Preferably, the GLAF attention mechanism includes global and local information extraction, adaptive fusion, direction-aware enhancement, and attention generation.

[0016] Preferably, the GLAF attention mechanism first uses global average pooling and multi-scale adaptive pooling to capture global context and local structural features, and enhance the spatial structure representation across multiple scales. Then, according to the input features, a lightweight dynamic weight module is used to adaptively adjust the fusion weights to achieve a more precise balance between global and local information, and a direction-aware enhancement strategy is used to improve spatial and directional consistency.

[0017] Preferably, the lightweight dynamic weight module uses input image features to calculate adaptive weights, automatically evaluates and adjusts the importance of global and local features, and allows GLAF to dynamically adjust its focus on relevant areas according to changes in input features. The lightweight dynamic weight module first compresses the input image feature information through two convolutional layers, then uses global average pooling to obtain significant feature information, and finally uses the Softmax function to assign adaptive weights.

[0018] Preferably, the direction-aware enhancement strategy extracts richer local features from multiple sampling directions. The direction-aware enhancement strategy uses convolutions of different sizes, including Conv 1×1 (·) is a convolution with a kernel of 1×1, which is used to preserve spatial details. 1×3 (·) and Conv 3×1 (·) is a convolution with kernels of 1×3 and 3×1, which effectively captures edge information in the horizontal and vertical directions, respectively.

[0019] Preferably, in step S2, the dynamic adaptive multi-scale feature fusion encoder first uses the self-attention mechanism to perform intra-scale interaction on the input multi-scale features {S3, S4, S5} to obtain multi-scale features {S3, S4, A5}, and then uses a bottom-up, top-down and horizontally connected multi-scale feature fusion method to obtain multi-scale features {F3, F4, F5}. Finally, the dynamic adaptive multi-scale feature fusion module is used to obtain the adaptively fused multi-scale features {D3, D4, D5}, which are input into the decoder.

[0020] Preferably, the dynamic adaptive multi-scale feature fusion module utilizes the attention guidance mechanism to dynamically adjust the fusion weight according to the relative importance of features of different scales. The dynamic adaptive multi-scale feature fusion module first scales the multi-scale features {F3, F4, F5} to match the resolution F3 to obtain {H3, H4, H5}, and then concatenates the scaled features along the channel dimension to form H = Cat ([H3, H4, H5]). The dynamic adaptive multi-scale feature fusion module inputs each multi-scale feature {H3, H4, H5} into the GLAF attention mechanism, generates an attention weight G for each feature map, and realizes global information exchange between channels through a strategy of redistributing attention weights between different scales, thereby obtaining the final distribution weight R of each feature channel, and finally connects the attention-guided weights in channel order, applies these weights to the feature H, and obtains the final fused feature.

[0021] The present invention also provides a lightweight real-time infrared target detection system for unmanned boats, comprising:

[0022] The image processing module is used to collect infrared video data of the unmanned boat surface environment, perform preprocessing, build an infrared image target data set, annotate and classify each target in the data set, and divide the annotated data set into a training set, a validation set, and a test set according to a certain ratio;

[0023] The model training module is used to train the lightweight real-time infrared target detection model through the data set to obtain a trained model. The detection process of the lightweight real-time infrared target detection model includes:

[0024] Using the lightweight backbone network GLAF-MobileNetV4 based on the global-local adaptive fusion attention mechanism, deep feature extraction is performed on infrared images to capture multi-scale and multi-level feature information in the image;

[0025] The multi-scale and multi-level feature information is fed into a dynamic adaptive multi-scale feature fusion encoder, and the minimum uncertainty query selection method is used to select key features from the encoder output as the initial object query;

[0026] The Transformer-based decoder refines these initial object queries to obtain target boxes and corresponding category confidences.

[0027] The model deployment module is used to convert the model weight file of the trained lightweight real-time infrared target detection model into the ONNX format first, and then further convert it into the model Engine supported by the TensorRT inference engine, and deploy it to the edge processing device of the unmanned boat to perform infrared target detection tasks in actual surface environments.

[0028] Beneficial effects of the present invention:

[0029] 1. In a complex and changeable marine environment, the present invention significantly improves the accuracy, stability and real-time performance of infrared sea surface target detection compared to advanced detection algorithms. In particular, the method and system proposed by the present invention can be integrated into an unmanned boat, which not only achieves an ultra-high recognition rate, but also makes its deployment process light and fast, greatly enhancing the overall detection efficiency.

[0030] 2. The detection method of the present invention adopts an end-to-end DETR method architecture. This innovative design omits the post-processing operation of the non-maximum suppression method, thereby effectively avoiding the instability and speed delay problems that may be caused by the traditional YOLO algorithm in the post-processing stage, and further improving the detection efficiency.

[0031] 3. The present invention proposes a global-local adaptive fusion attention mechanism. This mechanism can cleverly integrate global and local information, optimize the feature fusion process by dynamically adjusting weights, and thus significantly improve the detection performance of objects of different sizes. In addition, the design of this attention mechanism is lightweight and does not increase excessive computational costs.

[0032] 4. In terms of feature fusion, the present invention proposes a dynamic adaptive multi-scale feature fusion encoder. Different from the static fusion method, the encoder uses an attention guidance mechanism to dynamically adjust the fusion weights according to the relative importance of each scale feature. This adaptive method enables the model to process relevant features more efficiently and balance the contributions of small and large targets, thereby further optimizing the accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present invention or 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 in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 It is a schematic diagram of the lightweight real-time infrared target detection method and system composition for unmanned boats of the present invention;

[0035] Figure 2 This is a network structure diagram of a lightweight real-time infrared target detection method for unmanned boats of the present invention;

[0036] Figure 3 It is a structural diagram of the lightweight backbone network GLAF-MobileNetV4 based on global local adaptive fusion attention in an embodiment of the present invention;

[0037] Figure 4 is a structural diagram of global-local adaptive fusion attention in an embodiment of the present invention;

[0038] Figure 5 is a structural diagram of a dynamic adaptive multi-scale feature fusion module in an embodiment of the present invention;

[0039] Figure 6 2 is a comparison chart of the detection results in the embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0041] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0042] like Figure 1 As shown, the embodiment of this specification provides a lightweight real-time infrared target detection method for unmanned boats, comprising the following steps:

[0043] S1. Data collection and preprocessing: When the unmanned boat is operating in the actual water surface environment, the infrared thermal imager carried by it collects infrared video data of the water surface environment. Under the intelligent control of the unmanned boat main control board, these infrared video data are safely saved in a dedicated memory, providing a solid foundation for subsequent data processing and analysis. Subsequently, a series of carefully designed preprocessing processes are carried out on these infrared video data, including precise frame extraction operations, strict manual screening, efficient similarity image deduplication, and in-depth data cleaning, etc., to build a rich and high-quality infrared image target data set. In order to ensure the accuracy and practicality of the data set, professional annotation software is used to carry out detailed annotation and meticulous classification of each target in the data set. In the data division stage, a scientific random allocation method is adopted to strictly divide the annotated data into training set, validation set and test set according to the ratio of 8:1:1. In this division process, special attention is paid to maintaining the balance of the number of samples in each category in each sub-dataset to ensure that the model can fully learn the characteristics of various targets during the training process, so as to show excellent generalization ability and detection effect in practical applications.

[0044] In the specific implementation, in order to train a more robust and adaptable infrared target detection algorithm in step S1, the unmanned boat must ensure the diversity of scenes when collecting infrared data sets. The collection scenes covered by this method are extensive, including but not limited to various water environments such as inland lakes, rivers, offshore waters, coastlines, oceans, docks and ports. At the same time, the time period of data collection is also fully covered, from night to day, and then to dusk and dawn. In addition, considering the impact of weather conditions on infrared images, data collection is also carried out under various weather conditions such as clear, cloudy, rainy, foggy, big waves and snowy days. During the data collection process, the motion state of the unmanned boat also presents diversity, including stationary, uniform speed navigation, acceleration, deceleration and turning, etc. The changes in these motion states can further enrich the content of the data set and improve the adaptability of the detection algorithm to the dynamic environment of the unmanned boat. Before the formal data collection, the working parameters of the infrared thermal imager are accurately set according to specific needs, such as temperature measurement range, focal length and exposure time. The accurate setting of these parameters is crucial to ensure the quality and accuracy of the collected infrared images, and they will directly affect the effect of subsequent data processing and algorithm training. The types of sea surface targets covered in this dataset are also extremely rich, mainly including but not limited to liner, cargo ship, sailboat, canoe, container ship, fishing boat, yacht, speedboat, buoy, jet ski, paddle board, raft, fishery patrol boat, people, swimmers, etc. The comprehensive coverage of these categories will help improve the detection algorithm's ability to identify different types of sea surface targets, thereby showing stronger robustness and adaptability in practical applications.

[0045] During the specific implementation, in order to maximize the field of view and improve the observation effect in step S1, the infrared thermal imager is installed on the highest platform of the unmanned boat and equipped with an advanced pan-tilt system, which effectively realizes the anti-shake function and ensures that the collected images are stable and clear.

[0046] S2. Model design and training: The core components of the lightweight real-time infrared target detection method mainly include the lightweight backbone network GLAF-MobileNetV4, dynamic adaptive multi-scale feature fusion encoder and decoder. Its architecture is as follows Figure 2 As shown in Figure 1. The method achieves efficient and accurate infrared target detection through the following steps: First, the lightweight backbone network GLAF-MobileNetV4 is used, which incorporates the global local adaptive fusion attention mechanism to extract deep features from the input infrared image. This process can capture multi-scale and multi-level feature information in the image. Then, these multi-level feature information is fed into the dynamic adaptive multi-scale feature fusion encoder. In the encoder, the feature information not only fully interacts within the same scale, but also realizes dynamic fusion across scales, thereby effectively integrating complementary information from different levels and enhancing the feature expression ability. Subsequently, the minimum uncertainty query selection method is used to carefully select key features from the output of the encoder as the initial object query. Finally, the decoder based on the Transformer architecture refines these initial object queries. Through a complex self-attention mechanism and a feedforward neural network, the decoder gradually optimizes the object query, generates an accurate target box and the corresponding category confidence, and completes the infrared target detection task. In addition, the lightweight real-time infrared target detection method uses the infrared dataset obtained in step S1 for training and evaluation. The trained model weight file is first converted into ONNX (Open Neural Network Exchange) format, and then further converted into the model Engine supported by the TensorRT inference engine. This conversion process greatly simplifies the deployment process of the model on the edge computing device of the unmanned boat, providing strong technical support for the realization of real-time infrared target detection.

[0047] In the specific implementation, the lightweight real-time infrared target detection method in step S2 extracts target features at different levels from the infrared image through the lightweight backbone network GLAF-MobileNetV4 based on global local adaptive fusion attention. The lightweight backbone network GLAF-MobileNetV4 based on global local adaptive fusion (GLAF) attention is mainly divided into five stages, such as Figure 3As shown in the figure. Stage 1 usually contains fewer convolutional layers to capture basic spatial structure information from the original infrared image. As the network depth increases, stages 2 and 3 further extract more complex features. In particular, in order to enhance the expressiveness of features, the GLAF attention mechanism is cleverly integrated at the end of stages 2 and 3. The GLAF mechanism achieves accurate capture and enhancement of target features in infrared images by combining global context information and local detail features. Stages 4 and 5 constitute the core of the network, using a more complex combination of modules, including Extra Depthwise (ExtraDW), Inverted Bottleneck (IB), and the innovative GLAF-ConvNext module. The GLAF-ConvNext module embeds the GLAF mechanism, further improving the ability of feature fusion and extraction. It is worth noting that compared with stage 4, the convolution kernel size and number of channels in stage 5 have increased, aiming to capture richer feature details and wider context information. The ExtraDW module reduces the amount of computation through depthwise separable convolution while maintaining good feature extraction performance. The IB module adopts an inverted bottleneck structure, that is, the number of channels is expanded first and then the convolution operation is performed, and finally the number of channels is reduced by point-by-point convolution. This design effectively reduces the number of parameters and computational complexity while maintaining good feature expression capabilities. The GLAF-ConvNext module combines the advantages of the GLAF mechanism and the ConvNext module to achieve more efficient and powerful feature extraction. In addition, the ExtraDW, IB, and GLAF-ConvNext modules all adopt the UniversalInverted Bottleneck structure, which effectively expands the receptive field while improving computational efficiency by increasing the flexibility of spatial and channel mixing. GLAF-MobileNetV4 strategically reduces redundant block stacking and reduces the number of channels, thereby accelerating training and inference. By integrating the GLAF mechanism, it maintains a strong feature extraction capability for infrared maritime images. The feature information output by stages 3, 4, and 5 is used as the input of the encoder.

[0048] In specific implementation, the global local adaptive fusion GLAF attention mechanism in the lightweight backbone network GLAF-MobileNetV4 based on global local adaptive fusion attention in step S2 mainly consists of global and local information extraction, adaptive fusion, direction perception enhancement and attention generation, such as Figure 4As shown in the figure. The GLAF mechanism first applies global average pooling and multi-scale adaptive pooling to capture global context and local structural features, thereby enhancing the spatial structure representation across multiple scales. Then, the dynamic weight module calculates the adaptive fusion weights to balance the contribution of global and local information. In addition, the direction-aware enhancement module improves the spatial and directional consistency and improves the model's adaptability to objects of different shapes and sizes. This process ultimately generates effective fusion weights, thereby optimizing the overall feature extraction performance. The GLAF attention mechanism adopts a lightweight design, in which the adaptive average pooling operation is non-learning and has extremely low computational cost.

[0049] In this embodiment, the specific operation process of global-local adaptive fusion attention is as follows:

[0050] (1) Global average pooling is used to capture global context information. At the same time, multi-scale average pooling is used to extract local features. The specific operations are as follows:

[0051] L=[Scale(P(X,1)),Scale(P(X,2)),P(X,4)] T ,

[0052] Here, P(·,·) is the adaptive average pooling operation, and the numbers 1, 2, and 4 represent the target dimensions generated by the pooling method. Scale(·) is an operation that adjusts these outputs to a consistent dimension, thereby preserving spatial structural information and improving the comprehensive representation of diverse features at different scales.

[0053] (2) The GLAF attention mechanism adaptively adjusts the fusion weights according to the input features to achieve a more precise balance between global and local information. In order to generate adaptive weights, a lightweight dynamic weight module (DWM) is used. DWM automatically evaluates and adjusts the importance of global and local features as follows:

[0054] W=DWM(X),

[0055] Z=W T ·L,

[0056] in, is a learnable parameter vector.

[0057] (3) The fused features are enhanced with a direction-aware enhancement (DAE) strategy to improve the model’s adaptability to different target shapes and forms, and the output result is:

[0058] E = DAE(Z).

[0059] (4) The features are captured through the attention generation component to guide the model to focus on the most relevant areas and calculate the weight of each channel. Then, each channel of the input feature is multiplied by the corresponding weight to obtain the final output result:

[0060]

[0061] Where σ(·) represents the Sigmoid function; Represents channel-by-channel multiplication.

[0062] In specific implementation, the lightweight dynamic weight module DWM in the global-local adaptive fusion GLAF attention mechanism in step S2 uses the input image features to calculate adaptive weights, more accurately guiding the feature weights of different distances. DWM automatically evaluates and adjusts the importance of global and local features, allowing GLAF to dynamically adjust the focus on relevant areas according to changes in input features. DWM first compresses the input image features through two convolutional layers; then, global average pooling is used to obtain significant feature information, and finally the Softmax function is used to assign adaptive weights.

[0063] In specific implementation, the direction-aware enhanced DAE strategy in the global-local adaptive fusion GLAF attention mechanism in step S2 extracts richer local features from multiple sampling directions. Integrating these features can enhance the model's ability to capture multi-directional features of infrared maritime targets, thereby improving the detection performance of different object shapes and forms. The main operations of the direction-aware enhanced DAE are as follows:

[0064] E=Conv 1×1 (Conv 1×3 (Z)+Conv 1×1 (Z)+Conv 3×1 (Z)),

[0065] Among them, Conv 1×1 (·) is a convolution with a kernel of 1, which mainly retains spatial details. 1×3 (·) and Conv 3×1 (·) is a convolution with kernels of 1×3 and 3×1, which effectively captures edge information in the horizontal and vertical directions, respectively.

[0066] In the specific implementation, in step S2, the dynamic adaptive multi-scale feature fusion encoder first performs intra-scale interaction on the input multi-scale features {S3, S4, S5} to obtain multi-scale features {S3, S4, A5}; then, a bottom-up, top-down and horizontally connected multi-scale feature fusion method is used to obtain multi-scale features {F3, F4, F5}; finally, the dynamic adaptive multi-scale feature fusion module is used to obtain the output {D3, D4, D5}, as shown in Figure 2 shown.

[0067] In specific implementation, the dynamic adaptive multiscale feature fusion (DAMFF) module of the encoder in step S2 uses the attention guidance mechanism to dynamically adjust the fusion weight according to the relative importance of features of different scales. Through this adaptive method, the model can prioritize relevant features and balance the contributions of small and large targets, thereby optimizing detection accuracy. The DAMFF module is divided into three variants, DAMFF-1, DAMFF-2, and DAMFF-3, according to different output scales. The implementation process of each variant is the same. The detailed structure of DAMFF-1 is shown in the figure. Figure 5 As shown. The dynamic adaptive multi-scale feature fusion DAMFF module first scales the multi-scale features {F3, F4, F5} to match the resolution F3 to obtain {H3, H4, H5}. These scaled features are then concatenated along the channel dimension to form H = Cat([H3, H4, H5]). The DAMFF module inputs each multi-scale feature {H3, H4, H5} into the GLAF attention mechanism and generates an attention weight G for each feature map. This operation evaluates the importance of the relevant area within the feature. DAMFF achieves global information exchange between channels by redistributing the attention weight strategy between different scales, and obtains the final distribution weight R for each feature channel. The DAMFF module connects the attention-guided weights in channel order, and then applies these weights to the feature H to obtain the final fused feature.

[0068] S3. System Integration and Application: Deploy the trained model to the NVIDIA edge processing device of the unmanned boat. In order to realize the full range of functions of the unmanned boat, the target detection module is deeply integrated with other core functional modules (including but not limited to navigation, control, communication, etc.) to jointly build a fully functional and high-performance unmanned boat infrared target detection system. The unmanned boat system is applied to the actual surface environment to realize real-time infrared target detection tasks.

[0069] In specific implementation, in step S3, the unmanned boat uses its onboard infrared thermal imager to capture image data of the water surface environment in real time. These data are then sent to the NVIDIA edge processor, and the innovative lightweight real-time infrared target detection method is used to quickly and accurately detect the target in the image. Once the target is detected, the processor immediately sends the target information to the main controller of the unmanned boat. The main controller quickly judges and issues the next work instruction based on the received target information. These instructions may include tracking targets, obstacle avoidance operations, etc., to ensure that the unmanned boat can respond correctly according to the actual situation. At the same time, the main controller also sends the working status, location information, infrared monitoring images and other key data of the unmanned boat to the remote server in real time according to the relevant communication protocol through the data transmission and image transmission modules. The remote monitoring platform performs real-time data display and in-depth analysis by calling the data on the server. This not only provides an intuitive monitoring interface for operators, but also enables remote control and strategy adjustment of the unmanned boat based on real-time data, thereby ensuring that the unmanned boat can complete the infrared target detection task efficiently and safely.

[0070] This embodiment adopts data enhancement methods such as random color adjustment, random scaling, and random cropping, which not only enriches the infrared data set, but also improves the effect of infrared target detection. For network training, the parameters are set to batchsize 16, the number of iterations is 200, the Adam optimizer is used, the basic learning rate is 0.0001, and the weight decay is 0.0001.

[0071] This embodiment performs infrared target testing on the test set, and the test results are as follows: Figure 6As shown, the recognition results of this example are compared with DETR, YOLOv9, YOLOv10, and RT-DETR algorithms. This example can effectively detect infrared targets with high accuracy in scenes of similar objects, occlusion, targets of different sizes, and small targets at long distances. The specific actual detection performance comparison results are shown in Table 1. It can be clearly seen from the data in Table 1 that the embodiment of the present invention reduces the parameter amount and the calculation amount by 46.6% and 74.8% respectively compared with the YOLOv10 algorithm, and reduces by 60.7% and 71.1% respectively compared with the RT-DETR algorithm. In actual detection applications, the embodiment of the present invention achieves 146 FPS (frames per second), which is 67.8% higher than the YOLOv10 algorithm and 102.8% higher than the RT-DETR algorithm, which fully meets the high requirements of unmanned boats for real-time detection, and strongly proves that the present invention is a lightweight and efficient real-time infrared target detection method. Not only that, in terms of detection rate, the embodiment of the present invention also shows excellent performance. Under the condition of IOU (intersection over union) of 0.5, its average accuracy is improved by 10.1% compared with the YOLOv10 algorithm and 2.3% compared with the RT-DETR algorithm. In particular, the detection accuracy of small, medium or large targets is better than other existing detection methods, and fully meets the detection needs of unmanned boats in various scenarios.

[0072] Table 1 Comparison of detection performance of different detection algorithms for infrared targets

[0073]

[0074]

[0075] Those skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present invention is limited to these examples; under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A lightweight real-time infrared target detection method for unmanned boats, characterized in that: The following steps are involved: S1. Collect infrared video data of the unmanned boat surface environment, perform preprocessing, construct an infrared image target data set, annotate and classify each target in the data set, and divide the annotated data set into a training set, a validation set, and a test set according to a certain ratio; S2. Train the lightweight real-time infrared target detection model through the data set to obtain a trained model, wherein the detection process of the lightweight real-time infrared target detection model includes: Using the lightweight backbone network GLAF-MobileNetV4 based on the global-local adaptive fusion attention mechanism, deep feature extraction is performed on infrared images to capture multi-scale and multi-level feature information in the image; The multi-scale and multi-level feature information is fed into a dynamic adaptive multi-scale feature fusion encoder, and the minimum uncertainty query selection method is used to select key features from the encoder output as the initial object query; The Transformer-based decoder refines these initial object queries to obtain target boxes and corresponding category confidences. S3. Detect the target to be detected through the trained lightweight real-time infrared target detection model to obtain the detection result.

2. The lightweight real-time infrared target detection method for unmanned boats according to claim 1 is characterized in that: In step S1, the scenes of the collected infrared video data of the unmanned boat water surface environment include a variety of water environments, the time period of data collection covers a plurality of time periods, the weather conditions include different weather conditions, and the motion state of the unmanned boat includes a variety of motion states.

3. The lightweight real-time infrared target detection method for unmanned boats according to claim 1 is characterized in that: The lightweight backbone network GLAF-MobileNetV4 includes five stages. The first three stages include convolutional layers, and the GLAF attention mechanism is added at the end of the second and third stages. The fourth and fifth stages include Extra Depthwise, Inverted Bottleneck and GLAF-ConvNext modules, among which the GLAF-ConvNext module embeds the GLAF mechanism, and the Extra Depthwise, Inverted Bottleneck and GLAF-ConvNext modules all adopt the Universal Inverted Bottleneck structure.

4. The lightweight real-time infrared target detection method for unmanned boats according to claim 3 is characterized in that: The GLAF attention mechanism includes global and local information extraction, adaptive fusion, direction-aware enhancement, and attention generation.

5. The lightweight real-time infrared target detection method for unmanned boats according to claim 4 is characterized in that: The GLAF attention mechanism first uses global average pooling and multi-scale adaptive pooling to capture global context and local structural features and enhance the spatial structure representation across multiple scales. Then, according to the input features, a lightweight dynamic weight module is used to adaptively adjust the fusion weights to achieve a more precise balance between global and local information, and a direction-aware enhancement strategy is used to improve spatial and directional consistency.

6. The lightweight real-time infrared target detection method for unmanned boats according to claim 5 is characterized in that: The lightweight dynamic weight module uses input image features to calculate adaptive weights, automatically evaluates and adjusts the importance of global and local features, and allows GLAF to dynamically adjust its focus on relevant areas according to changes in input features. The lightweight dynamic weight module first compresses the input image feature information through two convolutional layers, then uses global average pooling to obtain significant feature information, and finally uses the Softmax function to assign adaptive weights.

7. The lightweight real-time infrared target detection method for unmanned boats according to claim 5 is characterized in that: The direction-aware enhancement strategy extracts richer local features from multiple sampling directions. The direction-aware enhancement strategy uses convolutions of different sizes, including Conv 1×1 (·) is a convolution with a kernel of 1×1, which is used to preserve spatial details. 1×3 (·) and Conv 3×1 (·) is a convolution with kernels of 1×3 and 3×1, which effectively captures edge information in the horizontal and vertical directions, respectively.

8. The lightweight real-time infrared target detection method for unmanned boats according to claim 1 is characterized in that: In step S2, the dynamic adaptive multi-scale feature fusion encoder first uses the self-attention mechanism to perform intra-scale interaction on the input multi-scale features {S3, S4, S5} to obtain the multi-scale features {S3, S4, A5}, and then uses the bottom-up, top-down and horizontally connected multi-scale feature fusion methods to obtain the multi-scale features {F3, F4, F5}. Finally, the dynamic adaptive multi-scale feature fusion module is used to obtain the adaptively fused multi-scale features {D3, D4, D5}, which are input into the decoder.

9. The lightweight real-time infrared target detection method for unmanned boats according to claim 8, characterized in that: The dynamic adaptive multi-scale feature fusion module uses the attention guidance mechanism to dynamically adjust the fusion weight according to the relative importance of features of different scales. The dynamic adaptive multi-scale feature fusion module first scales the multi-scale features {F3, F4, F5} to match the resolution F3 to obtain {H3, H4, H5}, and then splices the scaled features along the channel dimension to form H = Cat ([H3, H4, H5]). The dynamic adaptive multi-scale feature fusion module inputs each multi-scale feature {H3, H4, H5} into the GLAF attention mechanism, generates an attention weight G for each feature map, and realizes global information exchange between channels through the strategy of redistributing attention weights between different scales, thereby obtaining the final distribution weight R of each feature channel, and finally connects the attention-guided weights in channel order, applies these weights to the feature H, and obtains the final fused feature.

10. A lightweight real-time infrared target detection system for unmanned boats, characterized in that: include: The image processing module is used to collect infrared video data of the unmanned boat surface environment, perform preprocessing, build an infrared image target data set, annotate and classify each target in the data set, and divide the annotated data set into a training set, a validation set, and a test set according to a certain ratio; The model training module is used to train the lightweight real-time infrared target detection model through the data set to obtain a trained model. The detection process of the lightweight real-time infrared target detection model includes: Using the lightweight backbone network GLAF-MobileNetV4 based on the global-local adaptive fusion attention mechanism, deep feature extraction is performed on infrared images to capture multi-scale and multi-level feature information in the image; The multi-scale and multi-level feature information is fed into a dynamic adaptive multi-scale feature fusion encoder, and the minimum uncertainty query selection method is used to select key features from the encoder output as the initial object query; The Transformer-based decoder refines these initial object queries to obtain target boxes and corresponding category confidences. The model deployment module is used to convert the model weight file of the trained lightweight real-time infrared target detection model into the ONNX format first, and then further convert it into the model Engine supported by the TensorRT inference engine, and deploy it to the edge processing device of the unmanned boat to perform infrared target detection tasks in actual surface environments.

Images