Unmanned aerial vehicle maritime target dynamic identification method based on omnibearing feature fusion

Abstract

The invention discloses an unmanned aerial vehicle marine target dynamic identification method based on omnibearing feature fusion. The method comprises the following steps: A, obtaining an unmanned aerial vehicle image data set; b, constructing an unmanned aerial vehicle marine target dynamic identification model based on omnibearing feature fusion; c, training an unmanned aerial vehicle marine target dynamic identification model; d, optimizing the unmanned aerial vehicle marine target dynamic identification model; and E, performing target identification on the to-be-detected unmanned aerial vehicle image by using the optimized unmanned aerial vehicle marine target dynamic identification model to obtain a target identification result. According to the invention, the accuracy and robustness of target identification can be effectively improved, and the real boundary of the target can be better captured.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to a dynamic recognition method for unmanned aerial vehicle (UAV) maritime targets based on omnidirectional feature fusion in complex sea area scenarios. Background Art

[0002] In complex sea area scenarios, UAV maritime target recognition faces a series of severe challenges, such as sea surface reflection, light change, target occlusion, and dynamic changes of sea waves. These factors make it difficult to accurately capture the appearance, position, and contour of maritime targets, thus affecting the target recognition effect.

[0003] Currently, maritime target recognition methods generally adopt target detection technologies based on the YOLO series. For example, in the technical solution disclosed in the patent application with the application number 202411477253.8 and the patent name "A Fast Recognition Method for Maritime Ship Targets Based on YOLOv8": After collecting image samples of the ship navigation waters in a digital vehicle, the image samples are cleaned and labeled, and the image samples are enhanced to obtain a dataset that can be used for training and verification. A target recognition network model is constructed based on the YOLOv8 deep neural network. By setting hyperparameters such as the learning rate and batch size, and using a pre-trained model for transfer learning, the inference results of the model are post-processed by applying the non-maximum suppression algorithm to remove duplicate detection boxes and using a confidence threshold for result filtering to complete the prediction of ship targets. The trained model neural network is sparsely pruned and trained, and the pruned model is calibrated and quantified. The above technology reduces the memory consumption during device operation while maximizing the retention of the model detection accuracy, and at the same time speeds up the detection speed of the model.

[0004] Although the above-mentioned target detection technologies based on the YOLO series can provide good performance in real-time detection, these methods usually rely on rectangular boxes to label the position and size of targets, and cannot accurately outline the complex contours of targets. Especially when the target morphology is complex or partially occluded, the limitations of rectangular boxes are more obvious, reducing the accuracy and robustness of target recognition.

[0005] In the prior art, there are also target detection methods based on feature pyramid networks. For example, the technical solution disclosed in the patent application number 202311567359.2 and the patent name "Small Target Detection Algorithm for UAV Images Based on Sea Surface Rescue" adopts a scale selection pyramid network structure as the baseline network, and embeds a context enhancement module and a spatial attention module. Through the synergy of these modules, the algorithm can enhance the context feature information transmitted by the input image, deepen the feature layer's learning of relevant feature information, and improve detection accuracy. In the specific implementation, the model detector module adopts a balanced loss function, determines the optimal parameter value through experiments, and enhances the stability of the model during detection. Experiments show that the proposed algorithm can effectively improve the detection performance and recognition accuracy of drone images of sea targets, can reduce the time for sea rescue to find targets, and plays an extremely critical role in the success of rescue missions.

[0006] Although the above-mentioned target detection technology based on feature pyramid network can improve the robustness of target detection to a certain extent through multi-scale feature fusion and attention mechanism, it is not able to learn subtle pixel-level features in marine scenarios, which may lead to the attenuation of target edge information in the process of deep network transmission. In addition, the step-by-step feature reconstruction mechanism of the pyramid structure is prone to redundant calculations, making it difficult to achieve the requirements of low-power real-time detection. Summary of the invention

[0007] The purpose of the present invention is to provide a method for dynamic recognition of unmanned aerial vehicle maritime targets based on omnidirectional feature fusion, which can effectively improve the accuracy and robustness of target recognition and better capture the real boundary of the target.

[0008] The present invention adopts the following technical solutions:

[0009] A method for dynamic recognition of unmanned aerial vehicle maritime targets based on omnidirectional feature fusion includes the following steps:

[0010] A: Get the drone image dataset;

[0011] B: Construct a UAV maritime target dynamic recognition model based on all-round feature fusion. The UAV maritime target dynamic recognition model consists of a frequency domain enhancement and normalization module, a cross-graph coupling module, and an improved U-Net module based on dual-domain attention;

[0012] C: Using drone image datasets to train a drone marine target dynamic recognition model;

[0013] D: Optimizing the UAV maritime target dynamic recognition model using UAV image dataset;

[0014] E: Use the optimized UAV marine target dynamic recognition model to perform target recognition on the UAV image to obtain the target recognition result.

[0015] In step B, the frequency-domain enhancement and normalization module is used to perform frequency-domain enhancement and normalization on the preprocessed image, and through normalization, the image data is made to have a unified scale, and the normalized image is used as the input feature map of the cross-image coupling module; the cross-image coupling module is used to perform graph aggregation representation learning on the input feature map in the horizontal and vertical directions respectively, and couple the learning results in the two directions; the improved U-Net module based on dual-domain attention performs the target recognition task through the backbone network U-Net, combining the multi-layer spatial attention mechanism and the two-stream channel attention mechanism.

[0016] In step B, when performing frequency-domain enhancement on the image, the frequency-domain enhancement and normalization module first performs Fourier transform on each channel of the sea area image separately, converting the image from the spatial domain to the frequency domain; then filters and enhances the image in the frequency domain; finally, through inverse Fourier transform on each channel, the image is converted back from the frequency domain to the spatial domain.

[0017] In step B, when performing normalization on the image, the frequency-domain enhancement and normalization module first calculates the mean and standard deviation of each channel, and then subtracts the mean of each channel from each pixel value of each channel and divides by the standard deviation to ensure that the pixel values of the image fall within the set range.

[0018] In step B, the cross-image coupling module first performs graph aggregation learning in the horizontal direction on each channel of the input feature map, then sequentially recombines the row tensors after graph aggregation representation learning, restores each single-channel feature map, and then recombines all the single-channel feature maps into a horizontal-direction feature map according to the channel order; the cross-image coupling module performs the corresponding operations in the vertical direction of the input feature map according to the above method to obtain a vertical-direction feature map; finally, a coupling operation is performed on the feature maps in the horizontal and vertical directions.

[0019] In step B, the improved U-Net module based on dual-domain attention is adopted. By using the U-Net network as the backbone network, the multi-layer spatial attention mechanism is used before each downsampling stage to dynamically adjust the weights of each spatial position; at the same time, after splicing the encoder feature map and the decoder feature map in each skip connection, the two-stream channel attention mechanism is applied to dynamically adjust the weights of each channel in the horizontal and vertical directions, further optimizing the fusion effect of features at different levels.

[0020] In step B, for the improved U-Net module based on dual-domain attention, first, in a parallel dual-pooling manner, global average pooling and global max pooling are simultaneously performed on the input feature map in the channel dimension, and then the two are concatenated along the channel; subsequently, convolution operations are performed using convolution kernels of different sizes, and the obtained spatial weight maps containing information of different granularities are concatenated along the channel dimension and then calculated to obtain a spatial weight matrix; finally, the spatial weight matrix is multiplied by the input feature map to obtain a spatial attention feature map.

[0021] In step B, using the dual-stream channel attention mechanism, first, average pooling is respectively performed on the input feature map in the horizontal and vertical directions to extract the feature information of the sea area image from two complementary directions; then, the pooling results in the horizontal and vertical directions are concatenated in the spatial dimension, and a mixed attention map is obtained through a convolution operation; then, the mixed attention map is separated into horizontal and vertical components along the spatial direction, and the horizontal direction weight matrix and the vertical direction weight matrix are obtained respectively through convolution operations and activation functions on the two components; finally, the input feature map is multiplied by the horizontal direction weight matrix and the vertical direction weight matrix to obtain a dual-stream channel attention feature map.

[0022] In step C, the target dynamic recognition model is trained using the Adam optimizer, and the total loss is calculated through the weighted cross-entropy loss function formula.

[0023] The trained UAV sea target dynamic recognition model is tested for performance using precision, recall, F1-score, overall pixel accuracy, and intersection over union to optimize the target recognition model.

[0024] Aiming at the problems existing in UAV sea target recognition, such as sea surface reflection, illumination change, target occlusion, and dynamic change of sea waves, the present invention effectively enhances the details and contrast of the image, highlights the semantic regions of the targets in the sea area image, and at the same time ensures the consistency of the image data space, avoiding numerical deviations caused by different illuminations or sensor differences through frequency domain enhancement and normalization processing operations on the image. Then, by performing independent graph aggregation representation learning on the input feature map in the horizontal and vertical directions and coupling them, more abundant spatial information and local features are effectively captured. The present invention also classifies each pixel point using the improved U-Net network based on dual-domain attention to obtain the result of semantic segmentation, pays more attention to the features of important regions in the sea area image more accurately, improves the recognition ability of the targets, effectively suppresses the interference of the background region, highlights the feature expression of the target region, and finally completes the recognition of the UAV sea target, effectively improving the accuracy and robustness of UAV sea target recognition in complex sea area scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a schematic flowchart of the present invention;

[0026] Figure 2 This is the structural diagram of the dynamic recognition model for maritime targets of drones in the present invention;

[0027] Figure 3 This is the structural diagram of the cross - graph coupling module in the present invention;

[0028] Figure 4 This is the structural diagram of the multi - layer spatial attention mechanism in the present invention;

[0029] Figure 5 This is the structural diagram of the dual - stream channel attention mechanism in the present invention. Detailed implementation manners

[0030] The following describes the present invention in detail with reference to the accompanying drawings and embodiments:

[0031] As Figures 1 to 5 shown, the method for dynamically recognizing maritime targets of drones based on all - around feature fusion according to the present invention includes the following steps:

[0032] A: Obtain a drone image dataset, and construct a training set, a validation set, and a test set after pre - processing;

[0033] Among them, the drone image dataset is sourced from the OUC - UAV - SEG dataset; during pre - processing: First, perform data augmentation operations such as flipping and rotating the drone images in the dataset to expand the diversity and complexity of the dataset, enhance the adaptability of the model to images with different perspectives and directions, thereby improving the generalization ability of the model and effectively reducing overfitting. Subsequently, crop all the enhanced images to a unified predefined size to ensure that the input images have consistent spatial dimensions for convenient model training. Finally, divide the dataset into a training set, a validation set, and a test set according to a ratio.

[0034] In this embodiment, step A includes the following specific steps:

[0035] A1: Obtain a drone image dataset.

[0036] Obtain the OUC - UAV - SEG dataset. The OUC - UAV - SEG dataset uses 8 different types of drones to collect high - resolution images of the Bohai Sea coastal zone, including 726 original annotated images and 1750 unannotated images. The training set, validation set, and test set in the OUC - UAV - SEG dataset contain 477, 140, and 109 original images respectively. The OUC - UAV - SEG dataset contains a total of 15 semantic labels such as seagrass, enteromorpha, ocean, river, land, Spartina alterniflora, Suaeda salsa, vegetation, building, sky, road, ship, etc.

[0037] A2: Preprocess the UAV image dataset.

[0038] First, in order to overcome the problems of sparse dataset and single sample, and improve the robustness and generalization ability of the model, this embodiment uses a variety of data augmentation techniques for the UAV sea area images in the dataset. Specifically, for each sea area image, image flipping, rotation, scaling, and translation operations are required to expand the diversity of the dataset. These operations simulate the shooting scenarios of the UAV at different perspectives, headings, and flight altitudes, enhancing the model's adaptability to images in different directions, scales, and positions. Through these data augmentation methods, the model can learn image features more comprehensively, avoid overfitting, and then improve its generalization ability in complex environments, enabling the model to better handle target recognition tasks in unfamiliar scenarios.

[0039] Second, in order to ensure that the input images have consistent spatial dimensions and meet the input requirements of the model, this embodiment crops all the enhanced images to a unified size of 512x512. The cropping operation determines the appropriate cropping position according to the width and height of the image, ensuring that the targets in the image are not cropped off, especially larger or key targets. If the image size is small, the image is expanded by zero-padding to ensure that it meets the predefined size. Through this cropping step, the consistency of the input image size is guaranteed, the training efficiency of the model is enhanced, and its adaptability to targets of different sizes is improved.

[0040] A3: Divide the dataset after data augmentation and size adjustment into three parts: training set, validation set, and test set. The training set, validation set, and test set contain 3816, 1120, and 872 images respectively.

[0041] B: Construct a dynamic UAV maritime target recognition model based on all-round feature fusion;

[0042] In the present invention, the dynamic UAV maritime target recognition model based on all-round feature fusion consists of a frequency domain enhancement and normalization module, a cross-image coupling module, and an improved U-Net module based on dual-domain attention.

[0043] Among them, the frequency domain enhancement and normalization module is used to perform frequency domain enhancement and normalization processing on the preprocessed image, enhance the details and contrast of the image, and at the same time make the image data have a unified scale through normalization. The normalized image is used as the input feature map of the cross-image coupling module;

[0044] The cross-image coupling module is used to perform graph aggregation representation learning on the input feature map in the horizontal and vertical directions respectively, and couple the learning results in these two directions to capture richer spatial information and local features;

[0045] The improved U-Net module based on dual-domain attention uses the backbone network U-Net and combines a multi-layer spatial attention mechanism and a two-stream channel attention mechanism for target recognition tasks, effectively focusing on important regions and strengthening information interaction between different channels, thereby improving recognition accuracy and robustness.

[0046] In the dynamic recognition of unmanned aerial vehicle (UAV) maritime targets in complex sea area scenarios, the sea area images obtained by UAVs are usually affected by complex environmental factors, such as sea surface reflection, light changes, and sea wave dynamics. These factors make the target contours in the sea area images blurred and the details unclear. At the same time, due to factors such as sensor differences, light changes, and environmental noise, the pixel values of sea area images often have large fluctuations, resulting in possible large numerical deviations.

[0047] To address the above special problems in the dynamic recognition of UAV maritime targets in complex sea area scenarios, in the present invention, technical means of frequency domain enhancement processing and normalization processing are specially selected. First, the Fourier transform is performed on each channel of the image separately, then the contrast of the image is enhanced by adjusting the spectrum, and it is converted back to the spatial domain through the inverse Fourier transform. After frequency domain enhancement, the image is then normalized using the mean and standard deviation, that is, the pixel values of each channel of the UAV image are subtracted by the mean and divided by the standard deviation respectively. The normalized image is used as the input feature map of the cross-image coupling module. The above special processing can effectively enhance the details and contrast of the image, highlight the semantic regions of the targets in the sea area images, and at the same time ensure the consistency of the image data space, avoiding numerical deviations caused by different illuminations or sensor differences.

[0048] In the present invention, the specific steps of the operation of the frequency domain enhancement and normalization module are as follows:

[0049] a1: The frequency domain enhancement and normalization module performs frequency domain enhancement processing on the sea area image:

[0050] First, the Fourier transform is performed on each channel of the sea area image separately to convert the image from the spatial domain to the frequency domain; then, the image is filtered and enhanced in the frequency domain to suppress low-frequency noise and strengthen the high-frequency information in the image, so that the contrast between the maritime target and the background becomes more obvious, improving the resolution and recognition ability of the model for the target; finally, the inverse Fourier transform is performed on each channel to convert the image from the frequency domain back to the spatial domain.

[0051] a2: The frequency domain enhancement and normalization module performs normalization processing on the data space of the image:

[0052] First, calculate the mean and standard deviation of each channel. Then, subtract the mean of each channel from each pixel value of that channel and divide by the standard deviation to ensure that the pixel values of the image fall within the range of [-1, 1]. The above normalization operation can ensure that the feature distribution of the image is more uniform, accelerate the convergence of the model, and improve the robustness of the model under different environmental conditions.

[0053] In the dynamic recognition of maritime targets by drones in complex sea area scenarios, due to factors such as sea surface undulation, light reflection, and perspective transformation, the target features in sea area images often exhibit a highly irregular and complex spatial distribution, and usually show different feature patterns in the horizontal and vertical directions. In order to effectively extract the local features in these images and capture rich spatial information, a cross-graph coupling module is specially designed in the present invention to perform independent graph aggregation representation learning on the input feature map in the horizontal and vertical directions, and couple the learning results in the two directions, ultimately effectively enhancing the multi-dimensional spatial structure of sea area images and the reflection effect of target characteristics.

[0054] In the present invention, the specific operation steps of the cross-graph coupling module are as follows:

[0055] b1: The cross-graph coupling module first performs graph aggregation learning in the horizontal direction on each channel of the input feature map; that is, the cross-graph coupling module performs a row-by-row segmentation operation on each single-channel feature map, where each row tensor is regarded as a node in the graph structure. The connection rule of the nodes follows the principle of spatial proximity, that is, each node (representing a certain row) respectively establishes an edge connection with its directly adjacent nodes (that is, the row above and the row below), and a self-loop is added to each node. Then, apply graph convolution operation to update the features of these nodes, as shown below:

[0056]

[0057] Among them, is the updated feature of the p-th row, RELU represents the activation function, N p represents the neighbors of the p-th row, is the normalization constant, h q is the initial feature of the q-th row, W k is the trainable parameter;

[0058] b2: The cross-graph coupling module sequentially recombines the row tensors after graph aggregation representation learning to restore each single-channel feature map, and then reorganizes all the single-channel feature maps into a horizontal direction feature map according to the channel order. The horizontal direction feature map obtained by the above operation can effectively capture the context information along the horizontal axis in the sea area image, helping the model understand the horizontal distribution characteristics of the targets on the sea surface, such as the width and position relationship of the targets.

[0059] b3: According to the methods in steps b1 and b2, perform corresponding operations in the vertical direction of the input feature map to obtain the vertical direction feature map. Compared with the horizontal direction map aggregation, the vertical direction map aggregation focuses on the spatial information along the vertical axis, helping to identify the height changes and local details of the target, such as the depth features of the target and the impact of sea wave undulations on the target shape.

[0060] b4: In order to fully fuse the information in different spatial dimensions and thus capture more comprehensive and accurate target features, it is necessary to perform a coupling operation on the feature maps in the horizontal and vertical directions. Specifically, first, the horizontal direction feature map and the vertical direction feature map are concatenated according to the channel dimension, and then the original channel number is restored through 3 1*1 convolutional kernels to obtain the coupled feature map.

[0061] The specific calculation method is as follows:

[0062] OH = Conv 1*1 (Concat(S x ; S y ))))

[0063] Among them, OH represents the coupled feature map, Conv 1*1 is a 1*1 convolutional operation, Concat represents the concatenation operation, S x and S y represent the horizontal direction feature map and the vertical direction feature map respectively;

[0064] In the dynamic recognition of unmanned aerial vehicle (UAV) maritime targets in complex sea area scenarios, factors such as sea surface reflection, dynamic sea waves, and target occlusion will cause large changes in the position and shape of the target in the image, resulting in unreasonable weight allocation of the existing target dynamic recognition model in the image area and poor robustness and accuracy of segmentation. Therefore, in the present invention, an improved U-Net module based on dual-domain attention is adopted. By using the U-Net network as the backbone network and using a multi-layer spatial attention mechanism before each downsampling stage, the weights of each spatial position are dynamically adjusted, helping the model to automatically focus on more important regions when processing image features at different positions, thereby improving the robustness and accuracy of semantic segmentation. At the same time, in the present invention, after concatenating the encoder feature map and the decoder feature map in each skip connection, a two-stream channel attention mechanism is applied to dynamically adjust the weights of each channel from the horizontal and vertical directions, further optimizing the fusion effect of features at different levels, capturing the horizontal extensibility of the maritime target and the possible target height changes in the vertical direction, and enhancing the expression ability and recognition accuracy of the model.

[0065] In the present invention, in the improved U-Net module based on dual-domain attention, the multi-layer spatial attention mechanism can optimize the unique complexity in the sea area image. Especially when the shape of the target and the distribution of the background in space are very uneven, through multi-scale spatial weight mapping, the model can more accurately focus on the features of important regions in the sea area image and improve the target recognition ability. Its structural diagram is as shown in Figure 4 shown below.

[0066] d1: Due to the complexity of the sea area image, the image usually shows different texture and structural features in different regions. Relying solely on a single pooling method may not be able to comprehensively capture the rich semantic information in the image. Therefore, in this embodiment, in the way of parallel dual pooling, global average pooling and global max pooling in the channel dimension are respectively performed on the input feature map at the same time, and then the two are concatenated according to the channel:

[0067] a ave = Global_AvePooling(A)

[0068] a max = Global_MaxPooling(A)

[0069] a out = Concat(a ave ; a max )

[0070] Among them, a ave and a max respectively represent the results of global average pooling and global max pooling, a out represents the result after concatenating the two, Global_AvePooling and Global_MaxPooling respectively represent global average pooling and global max pooling, Concat represents the concatenation operation, and A represents the input feature map of the multi-layer spatial attention mechanism;

[0071] d2: The targets in the sea area image often have variable scales and irregular shapes. Especially in the case of wave undulation, dynamic background and partial occlusion of the target, the local features of the target show large differences at different scales. In order to effectively extract and fuse these multi-scale features, in this embodiment, through convolution operations with a filter size of 7*7 and a padding of 3, convolution operations with a filter size of 5*5 and a padding of 2, and convolution operations with a filter size of 3*3 and a padding of 1 respectively, three spatial weight mappings containing different granularity information are obtained. Then these three are concatenated according to the channel dimension to ensure that the model can not only perceive the large-scale features in the image but also capture the detailed local features. Then a spatial weight matrix is calculated using a 1*1 convolution kernel and a Sigmoid activation function.

[0072] d3: Multiply the spatial weight matrix by the input feature map to obtain the spatial attention feature map.

[0073] In the present invention, in the improved U-Net module based on dual-domain attention, the dual-stream channel attention mechanism can dynamically adjust the saliency of each feature point in two directions. It can not only optimize the weights for the characteristics of the target in different directions in the sea area image, but also effectively suppress the interference of the background area and highlight the feature expression of the target area. Its structural diagram is as Figure 5 shown.

[0074] The dual-stream channel attention mechanism includes the following:

[0075] e1: Perform average pooling on the input feature map in the horizontal and vertical directions respectively. The purpose of the above two average poolings is to extract the feature information of the sea area image from two complementary directions and provide a multi-perspective feature basis for the model. Among them, the horizontal direction focuses on capturing the horizontal structural features in the image, such as the wave texture in the sea area image or the targets distributed horizontally, while the vertical direction focuses on capturing the vertical information, such as the vertically arranged ships, buoys and other targets from the aerial perspective. The calculation method is as follows:

[0076] c x = AveragePooling x (F)

[0077] c y = AveragePooling y (F)

[0078] where c x and c y represent the pooling results in the horizontal and vertical directions respectively, AveragePooling x and AveragePooling y represent performing average pooling on the input feature map in the horizontal and vertical directions respectively, and F represents the input feature map of the dual-stream channel attention mechanism;

[0079] e2: In order to comprehensively integrate the feature distributions in the horizontal and vertical directions, further fuse and extract the mixed features, and enhance the collaborative expression ability of the model. In this embodiment, the pooling results in the horizontal and vertical directions are concatenated in the spatial dimension, and a mixed attention map is obtained through a 1*1 convolution operation:

[0080] Z = RELU(BN(Conv 1*1 (Concat(c x ; c y ))))

[0081] Among them, Z represents the mixed attention map, and BN is batch normalization.

[0082] e3: Separate the mixed attention map into horizontal and vertical components along the spatial direction, and apply a 1*1 convolution operation and a Sigmoid activation function to the two components respectively to obtain a horizontal direction weight matrix and a vertical direction weight matrix:

[0083] Z x ,Z y = Split(Z)

[0084] G x = Sigmoid(Conv 1*1 (Z x ))

[0085] G Y = Sigmoid(Conv 1*1 (Z y ))

[0086] Among them, Z x and Z y respectively represent the horizontal and vertical components separated from the mixed attention map, Split represents the separation operation, G x and G Y respectively represent the horizontal direction weight matrix and the vertical direction weight matrix; the horizontal direction weight matrix can highlight the horizontally extended ripple area, while the vertical direction weight matrix can enhance the vertically sparsely distributed target area, thereby improving the expression ability of image details and direction information.

[0087] e4: Multiply the input feature map by the horizontal direction weight matrix and the vertical direction weight matrix to obtain a two-stream channel attention feature map:

[0088]

[0089] Among them, F out represents the two-stream channel attention feature map, represents the multiplication operation.

[0090] C: Iteratively train the target dynamic recognition model using the training set until the number of training times reaches the preset number of times to obtain a trained target dynamic recognition model;

[0091] The training data during training is input into the target recognition model in batches for iterative training; and an optimizer, the number of iterations, the learning rate, the dropout rate, the batch size, and the loss function are set. In this embodiment, the number of iterations is set to 200, the learning rate is set to 0.0005, the batch size is set to 16, the optimizer is set to Adam, and the dropout rate is set to 0.1.

[0092] In order to learn the features of different categories more evenly and improve the generalization ability of the model, in this embodiment, the target dynamic recognition model is trained using the Adam optimizer, and the total loss is calculated through the weighted cross-entropy loss function formula:

[0093]

[0094] Where, represents the total loss; M corresponds to the total number of pixel points in a single image; T represents the number of target categories; β t represents the weight of category t. If category t in the dataset is a few-shot sample, then β t is set to 2, otherwise it is set to 1; b m,t is a binary indicator (0 or 1). If the true category of the m-th pixel point is t, then b m,t is equal to 1, otherwise b m,t is equal to 0; log is the logarithmic function, and v m,t represents the probability that the model predicts the category of the m-th pixel point as t.

[0095] In this embodiment, the model parameters obtained from each stage of training are saved, and at the same time, the validation dataset is used to verify and evaluate the target recognition model during the training process. After training is completed, the model with the best performance is selected from all the saved models as the trained target recognition model.

[0096] D: Use the test set to perform a performance test on the trained target recognition model to obtain the performance test results and optimize the target recognition model;

[0097] The test set is input into the trained target recognition model for testing. In this embodiment, according to the test results, the performance is evaluated using five evaluation metrics: precision (P), recall (R), F1 score, overall pixel accuracy (OA), and intersection over union (IoU).

[0098]

[0099] F1 = 2 * P * R / (P + R)

[0100]

[0101] Among them, TP represents the number of positive samples predicted as positive samples, FP represents the number of negative samples predicted as positive samples, FN represents the number of positive samples predicted as negative samples, TN represents the number of negative samples predicted as negative samples, and P, R, F1, OA, and IoU represent precision, recall, F1 score, overall pixel accuracy, and intersection over union respectively;

[0102] According to the performance test results, the trained target recognition model is optimized. Through repeated experiments and adjustments, based on the performance of the performance indicators, the model with the best performance is selected as the optimized target recognition model.

[0103] E: Using the optimized target recognition model, perform target recognition on the drone image to be detected, and obtain the results of target recognition.

[0104] The present invention proposes a dynamic recognition method for unmanned aerial vehicle (UAV) maritime targets based on omnidirectional feature fusion. First, perform frequency domain enhancement on the preprocessed sea area image to enhance the details and contrast of the target, and normalize the image. Then, let the feature map perform graph aggregation representation learning in the horizontal and vertical directions respectively, and couple the learning results in these two directions to capture richer spatial information and local features. Next, use an improved U-Net network based on dual-domain attention to classify each pixel point to obtain the result of semantic segmentation, and then complete the recognition of maritime targets by the UAV. Finally, perform performance testing on the trained neural network model. Through five evaluation indicators, namely precision (P), recall (R), F1 score, overall pixel accuracy (OA), and intersection over union (IoU), comprehensively evaluate the performance of the model on actual test data, and provide targeted guidance for the improvement of the model. In summary, the dynamic recognition method for UAV maritime targets based on omnidirectional feature fusion effectively improves the accuracy and robustness of UAV maritime target recognition in complex sea area scenarios.

[0105] The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and 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 various embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A dynamic recognition method for unmanned aerial vehicle (UAV) maritime targets based on omnidirectional feature fusion, characterized in that, It includes the following steps: A: Obtain the UAV image dataset; B: Construct a dynamic UAV maritime target recognition model based on all-round feature fusion. The dynamic UAV maritime target recognition model consists of a frequency domain enhancement and normalization module, a cross-map coupling module, and an improved U-Net module based on dual-domain attention; C: Use the UAV image dataset to train the dynamic UAV maritime target recognition model; D: Use the UAV image dataset to optimize the dynamic UAV maritime target recognition model; E: Use the optimized dynamic UAV maritime target recognition model to perform target recognition on the UAV image to be detected and obtain the target recognition result.

2. The dynamic recognition method of the unmanned aerial vehicle for maritime targets according to claim 1, characterized in that: In step B, the frequency domain enhancement and normalization module is used to perform frequency domain enhancement and normalization processing on the preprocessed image, and through normalization, the image data has a unified scale. The normalized image is used as the input feature map of the cross-map coupling module; The cross-map coupling module is used to perform graph aggregation representation learning on the input feature map in the horizontal and vertical directions respectively, and couple the learning results in the two directions; The improved U-Net module based on dual-domain attention performs the target recognition task through the backbone network U-Net, combined with a multi-layer spatial attention mechanism and a two-stream channel attention mechanism.

3. The method for dynamically identifying marine targets by an unmanned aerial vehicle according to claim 2, wherein: In step B, when performing frequency domain enhancement on the image, the frequency domain enhancement and normalization module first performs Fourier transform on each channel of the sea area image respectively, converting the image from the spatial domain to the frequency domain; then filters and enhances the image in the frequency domain; finally, through inverse Fourier transform on each channel, the image is converted back from the frequency domain to the spatial domain.

4. The method for dynamically identifying maritime targets of an unmanned aerial vehicle according to claim 2, wherein: In step B, when performing normalization processing on the image, the frequency domain enhancement and normalization module first calculates the mean and standard deviation of each channel, and then subtracts the mean of each channel from each pixel value of each channel and divides by the standard deviation to ensure that the pixel values of the image fall within the set range.

5. The dynamic identification method of the unmanned aerial vehicle for maritime targets according to claim 2, characterized in that: In step B, the cross-map coupling module first performs horizontal direction graph aggregation learning on each channel of the input feature map, then sequentially recombines the row tensors after graph aggregation representation learning, restores each single-channel feature map, and then recombines all the single-channel feature maps in the channel order into a horizontal direction feature map; The cross-map coupling module performs the corresponding operations in the vertical direction of the input feature map according to the above method to obtain a vertical direction feature map; Finally, a coupling operation is performed on the feature maps in the horizontal and vertical directions.

6. The method for dynamically identifying maritime targets of an unmanned aerial vehicle according to claim 2, wherein: In step B, an improved U-Net module based on dual-domain attention is adopted. By using the U-Net network as the backbone network, a multi-layer spatial attention mechanism is used before each downsampling stage to dynamically adjust the weights of each spatial position; At the same time, after splicing the encoder feature map and the decoder feature map in each skip connection, a two-stream channel attention mechanism is applied to dynamically adjust the weights of each channel in the horizontal and vertical directions, further optimizing the fusion effect of features at different levels.

7. The dynamic identification method of unmanned aerial vehicle for maritime targets according to claim 2, characterized in that: In step B, for the improved U-Net module based on dual-domain attention, first, in a parallel dual-pooling manner, global average pooling and global max pooling in the channel dimension are simultaneously performed on the input feature map, and then the two are concatenated along the channel; subsequently, convolution operations are carried out using convolution kernels of different sizes, and the obtained spatial weight maps containing information of different granularities are concatenated along the channel dimension and then the spatial weight matrix is calculated; finally, the spatial weight matrix is multiplied by the input feature map to obtain the spatial attention feature map.

8. The method for dynamically identifying maritime targets of an unmanned aerial vehicle according to claim 2, wherein: In step B, using the dual-stream channel attention mechanism, average pooling is first performed on the input feature map in the horizontal and vertical directions respectively to extract the feature information of the sea area image from two complementary directions; then the pooling results in the horizontal and vertical directions are concatenated in the spatial dimension, and a mixed attention map is obtained through convolution operations; then the mixed attention map is separated into horizontal and vertical components along the spatial direction, and the horizontal direction weight matrix and the vertical direction weight matrix are obtained respectively through convolution operations and activation functions for the two components; finally, the input feature map is multiplied by the horizontal direction weight matrix and the vertical direction weight matrix to obtain the dual-stream channel attention feature map.

9. The dynamic identification method of maritime targets for drones according to claim 1, characterized in that: In step C, the target dynamic recognition model is trained using the Adam optimizer, and the total loss is calculated through the weighted cross-entropy loss function formula.

10. The dynamic identification method of the unmanned aerial vehicle for maritime targets according to claim 1, wherein: The precision rate, recall rate, F1 score, overall pixel accuracy, and intersection over union are used to test the performance of the trained UAV maritime target dynamic recognition model to optimize the target recognition model.

Images