An infrared target detection method for water surface based on multi-scale feature adaptive fusion

Abstract

The application provides an infrared target detection method based on multi-scale feature adaptive fusion on water, which comprises the following steps: performing adaptive scaling on an input infrared image on water to unify the input image size; obtaining multi-scale feature information by using a feature extraction network; then, performing preliminary feature fusion on the multi-scale feature information by using a bidirectional jump connection feature fusion module; outputting final feature information by using the multi-scale feature adaptive fusion module after the preliminary fusion; wherein, the multi-scale feature adaptive fusion module adaptively adjusts the fusion proportion between different scale feature layers by using an efficient frequency channel attention module; finally, a prediction layer is used to predict the feature information to obtain multiple prediction boxes, a non-maximum suppression method is used to determine the target box, target category and confidence, so that the target is detected. The application can adapt to complex and changeable water scenes, and improve the accuracy and anti-interference ability of the infrared target detection on water.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method for detecting underwater infrared targets based on multi-scale feature adaptive fusion. Background Technology

[0002] Unmanned surface vessels (USVs) and unmanned aerial vehicles (UAVs) are indispensable key technologies, serving as primary means for future missions such as maritime surveillance, rescue, transportation, and air-sea coordination. They hold significant applications in both civilian and military fields. Maritime infrared target detection technology is a major research area within USVs, representing a core technology for achieving environmental perception in low-light and dark conditions. USVs can utilize this technology to achieve autonomous detection, decision-making, and obstacle avoidance in low-light environments.

[0003] Because the aquatic environment is more complex than the terrestrial environment, infrared target detection on water is often affected by heavy fog, waves, and sunlight reflection, resulting in numerous environmental interference factors that increase the difficulty of target detection. Furthermore, infrared targets lack color and texture information, resulting in weak feature information, making infrared target detection on water extremely challenging.

[0004] To address the aforementioned challenges in underwater infrared target detection, traditional algorithms typically design features based on image information such as the target's grayscale value and surrounding background information. These features, whether local or global, are manually designed feature models. Feature strategies are pre-set for different application scenarios, and once complex underwater infrared scenes exceed these pre-defined conditions, existing traditional algorithms become difficult to apply, hindering improvements in detection capabilities. With the rapid development of deep learning, convolutional neural network models have overcome the limitations of manually designed features and demonstrated excellent performance. Trained on datasets, they can autonomously design features, better adapting to underwater infrared target detection and improving robustness. Due to the unique characteristics of underwater infrared target images—often consisting of small targets lacking texture information—common deep learning-based infrared target detection methods employ simple feature fusion to enhance features. While this improves detection performance, it suffers from poor adaptability to complex underwater environments, low detection stability, and weak anti-interference capabilities, resulting in poor underwater infrared target detection performance and a high risk of misdetection. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a method for detecting underwater infrared targets based on multi-scale feature adaptive fusion, so as to solve the problem of poor target detection performance in complex aquatic environments.

[0006] To achieve the above objectives, this invention provides a method for detecting underwater infrared targets based on multi-scale feature adaptive fusion, comprising the following steps:

[0007] S1. Preprocess the input infrared image of the water surface;

[0008] S2. The preprocessed image is processed through a feature extraction network to obtain four different levels of multi-scale depth feature maps.

[0009] S3. The four different levels of multi-scale deep feature maps are fused through bidirectional skip connections to output the feature information of the four different scales after preliminary fusion.

[0010] S4. The feature information of the four different scales after the initial fusion is subjected to multi-scale feature adaptive fusion to output the final feature information. This includes obtaining the attention weights of different channels in the feature maps of different sizes through the high-efficiency frequency channel attention module, adaptively weighting the attention weights with the different channel information of the corresponding feature maps of different sizes, adaptively adjusting the fusion ratio between different feature layers, and outputting feature information of three scales.

[0011] S5. Input the feature information of the three scales into the prediction layer to predict multiple prediction boxes. Use the non-maximum suppression method to determine the final target box, target category and confidence from the multiple prediction boxes to obtain the target detection result.

[0012] Preferably, the preprocessing includes adaptive scaling of the input underwater infrared image, with a uniform input image size of 640×640×3.

[0013] Preferably, in step S2, the feature extraction network is a CSP-Darknet53 feature extraction network, including a CBS module, a C3 module, and an SPPF module. The CBS module is a convolutional module, consisting of Conv2d convolution, BatchNorm normalization, and SiLu activation function. The C3 module is a residual structure composed of three CBS convolutional modules. The SPPF module first passes through the CBS convolutional module, then through three 5×5 max pooling layers to extract features at different scales, and then achieves multi-scale feature fusion through channel merging.

[0014] Preferably, step S2 specifically includes the following sub-steps:

[0015] After inputting a 640×640×3 underwater infrared image into the feature extraction network, a 160×160×128 feature map P1 can be obtained through two CBS modules and one C3 module.

[0016] S22. After passing through a CBS module and a C3 module, an 80×80×256 feature map P2 is obtained. Then, after passing through a CBS module and a C3 module, a 40×40×512 feature map P3 is obtained. After passing through a CBS module, a C3 module and an SPPF module, a 20×20×1024 feature map P4 is obtained.

[0017] Preferably, step S3 further includes:

[0018] S31. Through bidirectional skip connection feature fusion, the input feature maps P1, P2, P3, and P4 are fused from deep features to shallow features using a combination of top-down and lateral skip connections, outputting features B1, B2, B3, and B4. The specific fusion process includes:

[0019] B1 = Conv(P1)

[0020] B1 is a horizontal jump connection from P1, where Conv is a convolution operation;

[0021]

[0022] B4 is formed by fusing the feature information of P4 and B3, where Upsample is the upsampling operation, and ω i The weights on each path are obtained through backpropagation. ε is the learning rate set to prevent numerical instability, with the parameter set to 0.0001. B3 is obtained from the feature information of P3 and B2 using the same fusion method, and B2 is obtained from the feature information of P2 and B1 using the same fusion method.

[0023] S32. Perform feature fusion on features B1, B2, B3, and B4 using low-to-high and lateral skip connections, outputting preliminary fused feature information F1, F2, F3, and F4 at four different scales. The specific fusion process includes:

[0024]

[0025] F1 is formed by fusing the feature information from B1, F2, and the laterally skipped P1 connection, where Downsample is the downsampling operation, and ω′ i The weights for each path are obtained through backpropagation. Following the same fusion method, F2 is formed by fusing the feature information from P2, B2, and F3, while F3 is formed by fusing the feature information from P3, B3, and F4.

[0026]

[0027] F4 is formed by fusing the feature information of P4 and B4.

[0028] Preferably, step S4 further includes the following sub-steps:

[0029] S41. The multi-scale feature adaptive fusion module takes the multi-scale F1, F2, F3, and F4 feature maps output by the feature extraction network and unifies them to the same scale and number of channels through scale changes. First, it adjusts according to the F1 scale, resulting in feature maps of the same scale: X1, X2, X3, and X4. Then, it concatenates the channels to unify the feature map into X.

[0030] X = Cat([X1,X2,X3,X4])

[0031] Cat(·) represents performing channel concatenation on all feature maps;

[0032] S42. Feature maps X obtained from different scale transformations i The attention weights E of different channels in different feature maps are obtained through a high-efficiency frequency channel attention module. i ,

[0033] E i =EFCA(X i ), i = 1, 2, 3, 4

[0034] Here, EFCA(·) represents using the efficient frequency channel attention module EFCA to obtain attention for different channels of different feature maps. i For feature X i Channel attention weight values;

[0035] S43. Perform channel splicing on the entire multi-scale channel attention to obtain a new weight E.

[0036]

[0037] in, Indicates channel splicing operation, E i Attention weights for each multi-scale channel;

[0038] S44. The channel attention E is adaptively redistributed using the normalized exponential function Softmax, thereby obtaining the multi-scale feature adaptively weighted channel attention weights att. i ,

[0039]

[0040] Where Softmax(·) represents the Softmax function, used to obtain the channel recalibration weights att in feature maps at different scales. i ;

[0041] S45, adjust the channel attention weight vector atti With feature maps X at different scales i The channel information is adaptively weighted, and multi-scale feature maps Y1 are output through parallel combination.

[0042]

[0043] in, The multiplication operator represents feature weighting. Channel concatenation operations can preserve feature information completely without destroying the original feature information;

[0044] S46. The multi-scale feature adaptive fusion module sequentially outputs three scale feature information Y1, Y2, and Y3 according to scales F1, F2, and F3, repeating steps S41-S45.

[0045] Preferably, step S42 specifically includes:

[0046] S421: EFCA compresses the size of the input feature map X, and then performs a discrete cosine transform to obtain an feature vector s = (s1, s2, ..., s...). c ), as the input to the one-dimensional convolution operation, where s = (s1, s2, ..., s) c The calculation method for ) is as follows:

[0047]

[0048] Where i∈{0,1,…,H-1} and j∈{0,1,…,W-1} represent the component indices of the DCT, and each block uses a different frequency component. These are the basic functions of DCT.

[0049]

[0050] S422: Information interaction between a channel and its k-1 neighboring channels is achieved through a one-dimensional convolution with kernel k, followed by a sigmoid function to obtain the weight vector E = (E1, E2, ..., E...). c ),

[0051] E=σ(C1D k (s))

[0052] Where σ is the Sigmoid function and the convolution kernel k = 5.

[0053] Preferably, in step S2, the feature extraction network needs to be trained before use. The training datasets are selected from the SMD dataset, VAIS dataset, and LSOTB-TIR dataset, which contain image data of infrared scenes on water. The datasets are randomly divided into training set, validation set, and test set, accounting for 80%, 10%, and 10% of the data, respectively.

[0054] Preferably, before network training, the initial anchor box size is automatically set according to different object detection datasets, the training parameters are set to batchsize of 16, the test batchsize is 1, the epoch is 500, the initial learning rate is 0.01, and the learning rate momentum is 0.937.

[0055] The beneficial effects of this invention are as follows: This invention is applicable to infrared target detection scenarios in unmanned aquatic systems. The aquatic infrared target detection method provided by this invention effectively solves the problems of poor detection performance and weak anti-interference capability in complex aquatic environments. Through the multi-scale feature adaptive fusion module combining high-efficiency frequency channel attention provided by this invention, important target information in infrared features is highlighted, aquatic infrared feature information is enhanced, and interference from complex aquatic environments and other irrelevant information is suppressed, thereby improving the accuracy and anti-interference capability of aquatic infrared target detection. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of the overall process of the target detection method of the present invention;

[0058] Figure 2 This is a general diagram of the network structure of the target detection method according to an embodiment of the present invention;

[0059] Figure 3 This is a detailed network structure diagram of the target detection method according to an embodiment of the present invention;

[0060] Figure 4 This is a structural diagram of the multi-scale feature adaptive fusion module according to an embodiment of the present invention;

[0061] Figure 5 This is a structural diagram of the high-efficiency frequency channel attention module according to an embodiment of the present invention;

[0062] Figure 6 This is a comparison chart of the detection results of an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0064] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0065] like Figure 1 and Figure 2 As shown in the embodiments of this specification, a method for detecting underwater infrared targets based on multi-scale feature adaptive fusion is provided, including the following steps:

[0066] S1. Preprocess the input water infrared image. Specifically, adaptively scale the input water infrared image to unify the input image size.

[0067] In practice, the water infrared image input in step S1 is acquired by an infrared camera from an unmanned water system. The acquired water infrared image is adaptively scaled to uniformly transform the input image size to 640×640×3.

[0068] S2. The preprocessed image is processed through a feature extraction network to obtain four different levels of multi-scale deep feature maps. The feature information of the four scales mainly includes target feature information of different scales: weak, small, medium and large.

[0069] In specific implementation, the input image in step S2 is processed by the CSP-Darknet53 feature extraction network to extract multi-scale depth feature maps at different levels. The CSP-Darknet53 feature extraction network mainly consists of a CBS module, a C3 module, and an SPPF module, as shown in the attached diagram. Figure 3As shown in the diagram, the CBS module is a basic convolutional module consisting of Conv2d convolution, BatchNorm normalization, and SiLu activation. The C3 module is a residual structure composed of three CBS convolutional modules. The feature map is first split into two parts according to the number of channels; one part is convolutionally processed, and the other part is used for residual construction. Finally, channels are merged. This module can alleviate the gradient vanishing problem and enhance the network's generalization ability. The SPPF module first uses the CBS convolutional module, then extracts features at different scales through three 5×5 max-pooling layers, and finally achieves multi-scale feature fusion through channel merging, improving the scale invariance of the feature extraction network.

[0070] In specific implementation, the four different levels of multi-scale depth feature maps in step S2 are the four feature maps with scales of 160×160, 80×80, 40×40, and 20×20 obtained by downsampling the original input image by 4 times, 8 times, 16 times, and 32 times.

[0071] In this embodiment, the specific process of step S2 includes:

[0072] S21: After a 640×640×3 underwater infrared image is input into the feature extraction network, a 160×160×128 feature map P1 can be obtained through two CBS modules and one C3 module.

[0073] S22: After passing through a CBS module and a C3 module, we obtain an 80×80×256 feature map P2. Then, after passing through a CBS module and a C3 module, we obtain a 40×40×512 feature map P3. Finally, after passing through a CBS module, a C3 module, and an SPPF module, we obtain a 20×20×1024 feature map P4.

[0074] S3. By fusing the four different levels of multi-scale depth feature maps through bidirectional skip connections, the feature information of the four different scales after preliminary fusion is output, which enriches the infrared feature information and enhances the detection capability of weak infrared targets.

[0075] In specific implementation, step S3 includes:

[0076] S31, the bidirectional skip connection feature fusion module mainly fuses feature information from different levels, enabling the network to continuously adjust weights to learn the importance of different input features, such as... Figure 3 As shown. Through bidirectional skip connection feature fusion, the input feature maps P1, P2, P3, and P4 are fused from deep features to shallow features using a combination of top-down and lateral skip connections, outputting features B1, B2, B3, and B4. The specific fusion process includes:

[0077] B1 = Conv(P1)

[0078] B1 is a horizontal jump connection from P1, where Conv is a convolution operation;

[0079]

[0080] B4 is formed by fusing the feature information of P4 and B3, where Upsample is the upsampling operation, and ω i The weights on each path are obtained through backpropagation. ε is the learning rate set to prevent numerical instability, with the parameter set to 0.0001. B3 is obtained from the feature information of P3 and B2 using the same fusion method, and B2 is obtained from the feature information of P2 and B1 using the same fusion method.

[0081] S32. Perform feature fusion on features B1, B2, B3, and B4 using low-to-high and lateral skip connections, outputting preliminary fused feature information F1, F2, F3, and F4 at four different scales. The specific fusion process includes:

[0082]

[0083] F1 is formed by fusing the feature information from B1, F2, and the laterally skipped P1 connection, where Downsample is the downsampling operation, and ω′ i The weights for each path are obtained through backpropagation. Following the same fusion method, F2 is formed by fusing the feature information from P2, B2, and F3, while F3 is formed by fusing the feature information from P3, B3, and F4.

[0084]

[0085] F4 is formed by fusing the feature information of P4 and B4.

[0086] S4. The feature information of the four different scales after the initial fusion is subjected to multi-scale feature adaptive fusion to output the final feature information. This includes obtaining the attention weights of different channels in the feature maps of different sizes through the high-efficiency frequency channel attention module, adaptively weighting the attention weights with the different channel information of the corresponding feature maps of different sizes, adaptively adjusting the fusion ratio between different feature layers, and outputting feature information of three scales.

[0087] In specific implementation, step S4 includes: In complex aquatic environments, to better highlight the feature information of small infrared targets and weaken complex background interference, a multi-scale feature adaptive fusion (MFAF) module is designed by combining an efficient frequency channel attention (EFCA) module and multi-scale feature information fusion. This module uses the attention module to adaptively weight multi-scale feature information for fusion, such as... Figure 4 As shown. Specifically:

[0088] S41. The multi-scale feature adaptive fusion module unifies the multi-scale F1, F2, F3, and F4 feature maps output by the feature extraction network to the same scale and number of channels through scale changes. Initially, it adjusts according to the F1 scale, resulting in feature maps of the same scale: X1, X2, X3, and X4. Then, it concatenates the channels to unify the feature map into X.

[0089] X = Cat([X1,X2,X3,X4])

[0090] Cat(·) represents performing channel concatenation on all feature maps;

[0091] S42. Feature maps X obtained from different scale transformations i The attention weights E of different channels in different feature maps are obtained through a high-efficiency frequency channel attention module. i ,

[0092] E i =EFCA(X i ), i = 1, 2, 3, 4

[0093] Here, EFCA(·) represents using the efficient frequency channel attention module EFCA to obtain attention for different channels of different feature maps. i For feature X i Channel attention weight values;

[0094] S43. Perform channel splicing on the entire multi-scale channel attention to obtain a new weight E.

[0095]

[0096] in, Indicates channel splicing operation, E i Attention weights for each multi-scale channel;

[0097] S44. The channel attention E is adaptively redistributed using the normalized exponential function Softmax, thereby obtaining the multi-scale feature adaptively weighted channel attention weights att. i ,

[0098]

[0099] Where Softmax(·) represents the Softmax function, used to obtain the channel recalibration weights att in feature maps at different scales. i ;

[0100] S45, adjust the channel attention weight vector att i With feature maps X at different scales i The channel information is adaptively weighted, and multi-scale feature maps Y1 are output through parallel combination.

[0101]

[0102] in, The multiplication operator represents feature weighting. Channel concatenation operations can preserve feature information completely without destroying the original feature information;

[0103] S46. The multi-scale feature adaptive fusion module sequentially outputs three scale feature information Y1, Y2, and Y3 according to scales F1, F2, and F3, repeating steps S41-S45.

[0104] This method not only takes into account multi-scale feature information, but also selectively processes key feature information, invests more attention weight on key areas of the target to obtain more detailed information of the target to be inspected, and relatively weakens the interference information of complex water environment.

[0105] In this embodiment, the specific implementation process of the high-efficiency frequency channel attention module in step S42 includes:

[0106] This example uses the High-Efficiency Frequency Channel Attention (EFCA) module as the core attention module. It weights the feature map based on the learned channel attention weights, assigning lower weights to features less relevant to the target and higher weights to those more relevant, thus highlighting important target information and suppressing other irrelevant information such as interference from the complex aquatic environment. Figure 5 As shown, the EFCA attention module uses discrete cosine transform to introduce more frequency components to fully utilize feature information, and appropriate cross-channel interaction can significantly reduce model parameters.

[0107] Step 1: EFCA compresses the input feature map X, and then performs a Discrete Cosine Transform (DCT) operation to obtain a feature vector s = (s1, s2, ..., s...). c ), which serves as the input for a one-dimensional convolution operation. Where s = (s1, s2, ..., s...). c The calculation method for ) is as follows:

[0108]

[0109] Where i∈{0,1,…,H-1} and j∈{0,1,…,W-1} represent the component indices of the DCT, and each block uses a different frequency component. These are the basic functions of DCT.

[0110]

[0111] Step 2: A one-dimensional convolution with kernel k is used to interact with the information of the channel and its k-1 neighboring channels, and then the weight vector E = (E1, E2, ..., E...) is obtained by passing the sigmoid function. c )

[0112] E=σ(C1D k (s))

[0113] Where σ is the Sigmoid function and the convolution kernel k = 5.

[0114] S5. Input the feature information of the three scales into the prediction layer to predict multiple prediction boxes. Use the non-maximum suppression method to determine the final target box, target category and confidence from the multiple prediction boxes to obtain the target detection result.

[0115] In the above embodiment, S1 is the input image to be detected, S2 is the target depth feature extraction process, S3 and S4 are multi-scale feature information fusion, and S5 predicts the feature maps at three scales to output the final detected target. In the actual process of underwater infrared target detection, the network model needs to be trained in advance to better complete the entire target detection process.

[0116] This embodiment selects image data containing underwater infrared scenes from the SMD, VAIS, and LSOTB-TIR datasets as the dataset. The datasets are randomly divided into training, validation, and test sets, accounting for 80%, 10%, and 10% of the data volume, respectively.

[0117] In this embodiment, due to the limited number of water infrared image datasets during training, random scaling, random cropping, and random arrangement of data augmentation methods are used to enrich the water infrared image dataset and improve the infrared target detection effect. Before network training, this embodiment automatically sets the initial anchor box size according to different target detection datasets. The training parameters in this embodiment are set to batchsize of 16, batchsize of 1 during testing, epoch of 500, initial learning rate of 0.01, and learning rate momentum of 0.937.

[0118] This embodiment performs tests on waterborne targets using a test set, and the test results are as follows: Figure 6 As shown, the left side displays the recognition results of the YOLOv5 algorithm, and the right side displays the recognition results of this example. Compared with the YOLOv5 algorithm, this example can effectively detect infrared targets on water under different backgrounds, weather conditions, and target sizes, with high accuracy.

[0119] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0120] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A water infrared target detection method based on multi-scale feature adaptive fusion, characterized in that, The method comprises the following steps: S1, preprocessing the input water infrared image; S2, obtaining four different levels of multi-scale deep feature maps through a feature extraction network from the preprocessed image; S3, fusing the four different levels of multi-scale deep feature maps through a bidirectional jump connection feature fusion to output four different scales of feature information after preliminary fusion; S4, performing multi-scale feature adaptive fusion on the four different scales of feature information after preliminary fusion to output final feature information, including obtaining attention weights of different channels in different size feature maps through an efficient frequency channel attention module, adaptively weighting the attention weights and different channel information of the corresponding different size feature maps, adaptively adjusting the fusion ratio between different feature layers, and outputting three scales of feature information; S5, inputting the three scales of feature information into a prediction layer to predict a plurality of prediction boxes, determining a final target box, a target class and a confidence from the plurality of prediction boxes through a non-maximum suppression method, and obtaining a target detection result; Step S3 further comprises: S31, fusing the input by bidirectional jump connection features, outputting the deep layer features to the shallow layer features by top-down and lateral jump combination mode 、 、 、 feature map, fusing the deep layer features to the shallow layer features by top-down and lateral jump combination mode 、 、 、 features, the specific fusion process includes: is derived from a transverse skip connection, where is a convolution operation; is fused by and feature information, in which, is an up-sampling operation, is a weight on each path, obtained by back propagation, is a learning rate set to prevent numerical instability, and the parameter is set to 0.0001, is fused by and feature information obtained according to the same fusion manner, is fused by and feature information obtained according to the same fusion manner; S32, to , , , Feature fusion of features jumping up from low to high and laterally, outputting four different scale feature information after preliminary fusion , , , Features, and the specific fusion process includes: is fused by , and the lateral jump connection feature information, wherein is a down-sampling operation, is a weight on each path, obtained by back propagation; according to the same fusion mode, is fused by , and feature information, is fused by , and feature information; is formed by and characteristic information fusion; Step S4 further comprises the following sub-steps: S41. The multi-scale feature adaptive fusion module adapts the multi-scale output of the feature extraction network. , , , Feature maps, through scale transformation, unify multi-scale feature maps to the same scale and number of channels, the first time according to The scale is adjusted to obtain feature maps of the same scale. , , and Then, the unified feature map is obtained by channel splicing. , wherein, denotes a channel concatenation operation on all feature maps; S42, feature maps transformed by different scales Obtain different channel attention weights in different feature maps through the high-efficiency frequency channel attention module , wherein, representing using an efficient frequency channel attention module EFCA to obtain attention of different channels of different feature maps; for the feature channel attention weight value; S43, channel splicing is performed on the whole multi-scale channel attention to obtain new weights , wherein, represents a channel concatenation operation, is the multi-scale channel attention weight for each; S44, re-adapting the channel attention by using a normalized exponential function Softmax , thereby obtaining the channel attention weights of the multi-scale feature adaptive weighting , ， wherein, denotes a Softmax function for obtaining channel rescaling weights in different scale feature maps ; S45, the channel attention weight vector with channel information of different scale feature maps is adaptively weighted, and a multi-scale feature map is output by parallel combination , wherein, represents a multiplication operator symbol of feature weighting; the channel concatenation operation can completely maintain the feature information without destroying the original feature information; S46, the multi-scale feature adaptive fusion module sequentially according to 、 、 Scale, repeat the steps of S41-S45 to output three scale feature information 、 、 .

2. The method of claim 1, wherein the method comprises: The preprocessing comprises adaptively scaling the input water infrared image, and uniformly setting the input image size to 640x640x3.

3. The method of claim 2, wherein the method comprises: In step S2, the feature extraction network is a CSP-Darknet53 feature extraction network, which comprises a CBS module, a C3 module and a SPPF module. The CBS module is a convolution module composed of a Conv2d convolution, a BatchNorm standardization and a SiLu activation function. The C3 module is a residual structure composed of three CBS convolution modules. The SPPF module first passes through a CBS convolution module, then passes through three 5x5 maximum pooling layers to extract different scale features, and then realizes multi-scale feature fusion through channel merging.

4. The method of claim 3, wherein the method comprises: Step S2 specifically comprises the following sub-steps: S21, after the 640x640x3 water infrared image input feature extraction network, two CBS modules and a C3 module can obtain a 160x160x128 feature map ; S22, through a CBS module and a C3 module, a feature map of 80x80x256 is obtained , and then through a CBS module and a C3 module, a feature map of 40x40x512 is obtained , and then through a CBS module, a C3 module and an SPPF module, a feature map of 20x20x1024 is obtained .

5. The method of claim 1, wherein the method comprises: Step S42 specifically comprises: S421: EFCA on input feature map performing size compression, obtaining a feature vector through a discrete cosine transform operation , as an input of one-dimensional convolution operation, wherein is calculated in the following manner: wherein and denotes the component index of the DCT, each block employing a different frequency component, , is the basic function of the DCT, S422: pass through the one-dimensional convolution of the channel and its adjacent channel information to interact information, and then get the weight vector through the Sigmoid function , wherein is a sigmoid function, a convolution kernel .

6. The method of claim 1, wherein the method comprises: In step S2, the feature extraction network needs to be trained before use. The images containing water infrared scenes in the SMD dataset, the VAIS dataset and the LSOTB-TIR dataset are selected as the dataset. The dataset is randomly divided into a training set, a validation set and a test set, and the data amount of each set is 80%, 10% and 10% respectively.

7. The method of claim 6, wherein the method comprises: Before network training, the size of the initial anchor box is automatically set according to different target detection datasets. The training parameters are set as follows: batchsize is 16, batchsize is 1 during testing, epoch is 500, initial learning rate is 0.01, and learning rate momentum is 0.937.

Images