A method for underwater infrared target detection based on dense bidirectional pyramid network

Abstract

The present invention relates to the field of infrared target detection technology, and in particular to a method for detecting infrared targets on water based on a dense bidirectional pyramid network. The method comprises adaptively scaling a captured infrared image on water to a uniform size, inputting the image into a feature extraction network to extract multi-scale depth feature maps at different levels, inputting the multi-scale feature maps at different levels into a dense bidirectional pyramid network to fuse the multi-scale features, and adding an adaptive dual-squeeze excitation channel attention mechanism before outputting the multi-scale fused features. The multi-scale fused features are respectively input into a classification prediction network and a prediction box regression network to obtain the position and category information of all targets in the infrared image on water. The method enriches infrared target feature information, highlights important target information in the features, and can effectively improve the accuracy and stability of infrared target detection on water.

Description

Technical Field

[0001] The present invention relates to the field of infrared target detection technology, and in particular to a method for detecting infrared targets on water based on a dense bidirectional pyramid network. Background Art

[0002] With the intensification of globalization, the rapid development of maritime trade, and the increasing number of ship accidents, there is an urgent need to improve maritime monitoring and search and rescue technologies. The air-sea cross-domain autonomous collaborative system for maritime search and rescue is a heterogeneous multi-agent system composed of unmanned boats and drones. It can effectively improve the efficiency of maritime monitoring and search and rescue, reducing search and rescue costs and personnel risks. Overwater infrared target detection technology is one of the key technologies for rapidly monitoring and searching for targets awaiting rescue on the water under low-light conditions.

[0003] To achieve rapid monitoring and rescue, targets must be detected at a considerable distance. This results in infrared targets occupying a very limited space within the monitoring field of view, and most of these targets are small. Due to their small imaging area and low resolution, small targets carry little characteristic information, increasing the difficulty of detecting infrared targets above water. Furthermore, infrared targets lack color and texture information and have low contrast, further complicating the detection task.

[0004] To address the aforementioned difficulties in detecting infrared targets above water, traditional algorithms typically design features based on relevant image information, such as the target's grayscale value and surrounding background information, for detection and recognition. These features, whether local or global, are all manually designed feature models. However, once small infrared targets above water exceed preset conditions, existing algorithms that pre-set feature strategies based on different scenarios become difficult to apply, and their detection capabilities are therefore limited. With the rapid development of deep learning, convolutional neural network models have overcome the limitations of manually designed features and demonstrated excellent performance. By enabling autonomous feature design through dataset training, convolutional neural network models are better suited for infrared target detection and enhance robustness. Common deep learning-based infrared target detection methods use simple feature fusion to enhance infrared features, resulting in improved detection performance. However, due to the unique characteristics of infrared target images above water, which are often small and weak and lack texture information, direct application of deep learning-based infrared target detection methods to water targets cannot guarantee detection accuracy and stability. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to propose an overwater infrared target detection method based on a dense bidirectional pyramid network to solve the problem of poor detection effect of small infrared targets in complex overwater environments in existing detection methods.

[0006] Based on the above objectives, the present invention provides a method for detecting infrared targets on water based on a dense bidirectional pyramid network, comprising the following steps:

[0007] S1, adaptively scaling the collected water infrared images to a uniform size;

[0008] S2. Input the image into the feature extraction network to extract multi-scale depth feature maps at different levels;

[0009] S3. Inputting the obtained multi-scale feature maps of various levels into a dense bidirectional pyramid network to fuse the multi-scale features. The dense bidirectional pyramid network adds an adaptive dual-squeeze excitation channel attention mechanism before outputting the multi-scale fusion features.

[0010] S4. The multi-scale fusion features output by the dense bidirectional pyramid network are input into the classification prediction network and the prediction box regression network respectively to obtain the position and category information of all targets in the water infrared image.

[0011] Preferably, the feature extraction network is an improved EfficientNetV2 network, which includes 7 stages. The feature maps output by the third stage, the fourth stage, the fifth stage, and the sixth stage, as well as the feature map of the sixth stage that is downsampled once, are selected as the input of the dense bidirectional pyramid network.

[0012] Preferably, the improved EfficientNetV2 network includes a Fused-MBConv module and an MBConv module. The Fused-MBConv module is a residual network structure, and the residual part consists of a 3×3 ordinary convolution, a 3×3 depth-separable convolution, a residual structure of the SE channel attention mechanism, a 1×1 ordinary convolution, and a Dropout layer.

[0013] The SE channel attention mechanism consists of a global average pooling, two fully connected layers and a Sigmoid activation function.

[0014] Preferably, the improved EfficientNetV2 network consists of 7 stages. The first stage is a 3×3 ordinary convolution with a stride of 2, including BN and SiLu, and the output is the P1 layer.

[0015] The second to fifth stages are repeated stacking of Fused-MBConv modules. The second stage repeats the Fused-MBConv module twice with a step size of 1. The third stage repeats the Fused-MBConv module four times with a step size of 2, and the output is the P2 layer. The fourth stage repeats the Fused-MBConv module four times with a step size of 2, and the output is the P3 layer. The fifth stage repeats the Fused-MBConv module six times with a step size of 2, and the output is the P4 layer.

[0016] The sixth stage repeats the MBConv module 9 times with a step size of 2, and the output is the P5 layer;

[0017] The seventh stage consists of a normal 1×1 convolutional layer, an average pooling layer, and a fully connected layer;

[0018] The P6 layer is obtained by downsampling the P5 layer once;

[0019] The input of the dense bidirectional pyramid network is the feature information of the P2 layer, P3 layer, P4 layer, P5 layer and P6 layer extracted by the feature extraction network.

[0020] Preferably, in step S3, the fusion of multi-scale features adopts a multi-scale feature fusion method using top-down, bottom-up and jump dense connections, and the dense bidirectional feature pyramid network adds different normalized weights to each node input feature.

[0021] Preferably, the jump dense connection adopts the interval jump connection mode, and the first node receives the data from the previous node with a depth of 2. n The sum of the characteristic information of the distance, where l≥2, n≥0, 2 n <l.

[0022] Preferably, the normalized weight method is described as follows:

[0023]

[0024] Among them, I i is the input feature, O is the output fusion feature, ω i and ω j is a learnable weight, which is limited to between 0 and 1, ε is the learning rate, which is 0.001, and i and j are the number of network layers.

[0025] Preferably, in step S3, the adaptive dual-squeeze excitation channel attention mechanism includes:

[0026] Use global average pooling and global standard deviation pooling to squeeze the feature information into channels and obtain statistical information of each channel;

[0027] Use a fully connected layer and activation function ReLU to perform channel compression, and a fully connected layer to restore the number of channels to obtain relevant information between channels;

[0028] The attention weights of each channel are obtained after nonlinear normalization using the H-Sigmoid activation function;

[0029] The attention weight of each channel is multiplied by the feature map, and a self-learned weight factor is added to balance the information loss caused by attention, and then added to the original feature map.

[0030] Preferably, the specific process of the dual-squeeze excitation channel attention mechanism includes:

[0031] make The input feature is of size C×H×W. The feature is subjected to global average pooling and global standard deviation pooling respectively, and the result is

[0032]

[0033]

[0034] Perform channel splicing on the above features to obtain features t c =Cat([μ c ,σ c ]), where Cat(·) represents the channel concatenation operation on all feature maps;

[0035] Use the excitation operation to reduce the amount of computation and obtain the channel attention map, including fully connecting the feature T with parameter W0 to obtain a feature of dimension C / 2×1×1;

[0036] Use the ReLU activation function δ and restore the original channel dimension C×1×1 of the image input through a fully connected layer with parameter W1;

[0037] After normalization using the H-Sigmoid activation function, the attention weights S of each channel are obtained as follows:

[0038] S=σ(W1δ(W0T))

[0039] The attention weight S of each channel is multiplied by the feature map, and then added to the original feature map after adding the self-learning weight factor α to obtain the following output features

[0040] E=α(SX)+X

[0041] Among them, the weight factor α is a learnable parameter.

[0042] Preferably, in step S4, the classification prediction network and the prediction box regression network use the fusion features output by the dense bidirectional feature pyramid network to respectively predict the category and bounding box of the above-water infrared target, and the two networks share the same feature network weights;

[0043] Sorting and non-maximum suppression are performed on the predicted target results to filter out the boxes and scores of each category whose scores are greater than the confidence threshold;

[0044] Non-maximum suppression is performed based on the position and score of the frame to obtain the position and category information of all targets in the water infrared image.

[0045] The present invention has the following beneficial effects: It is applicable to infrared target detection on water in air-sea cross-domain water surveillance and search and rescue scenarios. The infrared target detection method provided by the present invention effectively solves the problem of poor detection of small infrared targets on water in complex water environments. The dense bidirectional pyramid network provided by the present invention adds more feature information, thereby enriching infrared target feature information, reducing the risk of gradient vanishing, and improving network training effects. The adaptive dual-squeeze excitation channel attention mechanism highlights important target information in the features. The present invention can effectively improve the accuracy and stability of infrared target detection on water. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 It is a schematic diagram of the overall process of the target detection method of the present invention;

[0048] Figure 2 This is the overall network structure diagram of the target detection method of the present invention;

[0049] Figure 3 2 is a structural diagram of a Fused-MBConv module according to an embodiment of the present invention;

[0050] Figure 4 is a structural diagram of the MBConv module according to an embodiment of the present invention;

[0051] Figure 5 This is a structural diagram of the adaptive dual-squeeze excitation channel attention mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] 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 with reference to specific embodiments.

[0053] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings 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 object being described changes, the relative positional relationship may also change accordingly.

[0054] like Figure 1 As shown, the embodiment of this specification provides a method for detecting infrared targets on water based on a dense bidirectional pyramid network. Figure 1 This is a schematic diagram of the overall process of the target detection method of the present invention. Figure 1 As shown, the following steps are included:

[0055] S1: The collected water infrared images are adaptively scaled to a uniform size;

[0056] S2: Input the image into the feature extraction network to obtain 5 different levels of multi-scale depth feature maps.

[0057] In specific implementation, the water infrared image input in step S1 is collected by an unmanned water system using an infrared thermal imager, and the collected water infrared image is adaptively scaled to transform the input image size into a uniform size.

[0058] In specific implementation, the water infrared image input in step S2 is used to extract deep feature maps at different levels through a feature extraction network. The feature extraction network is an improved EfficientNetV2 network, mainly composed of a Fused-MBConv module and an MBConv module. The network is divided into seven stages in total. The feature maps output by the third, fourth, fifth, and sixth stages, as well as the feature map of the sixth stage that has been downsampled once, are selected as the input of the dense bidirectional pyramid network, as shown in the following example. Figure 2 shown.

[0059] In the improved EfficientNetV2 network, the Fused-MBConv module is a residual network structure, such as Figure 3As shown in the figure, the residual part consists of a 3×3 ordinary convolution (including batch normalization (BN) and activation function SiLu), a 1×1 ordinary convolution (including BN), and a Dropout layer. When the stride is 2, the 3×3 convolution is directly connected to the 1×1 convolution; when the stride is 1, the 1×1 convolution is skipped and the 3×3 convolution is directly connected to the Dropout layer.

[0060] In the improved EfficientNetV2 network, the MBConv module is a residual network structure, such as Figure 4 As shown in the figure, the residual part consists of a 1×1 ordinary convolution (including BN and SiLu), a 3×3 depth-wise separable convolution (including BN and SiLu), the residual structure of the SE channel attention mechanism, a 1×1 ordinary convolution (including BN), and a dropout layer. The SE channel attention mechanism consists of a global average pooling, two fully connected layers, and a sigmoid activation function.

[0061] In this embodiment, the specific process of the feature extraction network in step S2 includes:

[0062] (1) The improved EfficientNetV2 network is divided into 7 stages in total. The first stage is a 3×3 ordinary convolution with a step size of 2, including BN and SiLu, and the output is the P1 layer.

[0063] (2) The second to fifth stages are repeated stacking of Fused-MBConv modules; the second stage repeats the Fused-MBConv module twice with a step size of 1; the third stage repeats the Fused-MBConv module four times with a step size of 2, and the output is the P2 layer; the fourth stage repeats the Fused-MBConv module four times with a step size of 2, and the output is the P3 layer; the fifth stage repeats the Fused-MBConv module six times with a step size of 2, and the output is the P4 layer.

[0064] (3) In the sixth stage, the MBConv module is repeated 9 times with a step size of 2, and the output is the P5 layer.

[0065] (4) The seventh stage consists of a normal 1×1 convolutional layer (including BN and SiLu), an average pooling layer and a fully connected layer.

[0066] (5) The P6 layer is obtained by downsampling the P5 layer once; and the feature information of the P2 layer, P3 layer, P4 layer, P5 layer and P6 layer extracted by the feature extraction network is sent to the dense bidirectional pyramid network.

[0067] S3: Input the multi-scale feature maps of five different levels into a dense bidirectional pyramid network to fuse the multi-scale features; the dense bidirectional pyramid network adds an adaptive dual-squeeze excitation channel attention mechanism before outputting the multi-scale fusion features.

[0068] In specific implementation, the dense bidirectional feature pyramid network in step S3 adopts a multi-scale feature fusion method of top-down, bottom-up and jump dense connection, such as Figure 2 As shown in the figure, the current node receives the sum of feature information from top-down, bottom-up, and jump dense connections to add more feature fusion and enrich the infrared feature information. To reasonably distribute the sum of input feature information of different resolutions, the dense bidirectional feature pyramid network adds different normalization weights to each node input feature. The fast weight normalization method is described as follows:

[0069]

[0070] Among them, I i is the input feature, O is the output fusion feature, ω i and ω j is a learnable weight, which is limited to between 0 and 1, ε is the learning rate, which is 0.001, and i and j are the number of network layers.

[0071] In specific implementation, the jump dense connection method of the dense bidirectional feature pyramid network in step S3 is a connection method based on interval jumps, such as Figure 2 As shown; the first node receives the data from the previous node with a depth of 2 n The sum of the characteristic information of the distance, where l≥2, n≥0, 2 n This skip connection method can effectively slow down the phenomenon of gradient vanishing, allowing the deep network to retain subtle information in the shallow network, avoiding the increase in computational complexity and information redundancy caused by additional connections.

[0072] In this embodiment, the specific fusion process of the dense bidirectional feature pyramid in step S3 is as follows: Figure 2 As shown in the figure, taking the 6th and 5th layers as examples, the fusion method of each node is described as follows. The fusion method of the 4th and 3rd layers in the dense bidirectional feature pyramid is similar to that of the 5th layer, and the fusion method of the 2nd layer is similar to that of the 6th layer.

[0073] The fusion method of the 6th layer is described as:

[0074]

[0075]

[0076]

[0077] Among them, Conv is the convolution operation, Downsample is the downsampling operation, ω i is the weight on each path, ε is 0.001, P6 in The input of the 6th layer feature, P5 m2 、P5 m4 For the 5th layer intermediate node, P6 m2 、P6 m4 For the 6th layer intermediate node, P5 out 、P6 out It is the feature output of layer 5 and layer 6. It should be noted that the feature output of layer 6 enters the adaptive dual-squeeze excitation channel attention mechanism.

[0078] The fusion method of the 5th layer is described as:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] Among them, Conv is the convolution operation, Upsample is the upsampling operation, Downsample is the downsampling operation, ω i is the weight on each path, ε is 0.001, P5 in 、P6 in For the feature input of the 5th and 6th layers, P4 m2 、P4 m4 For the 4th layer intermediate node, P5 m1 、P5 m2 、P5 m3 、P5 m4 、P5 m5 For the 5th layer intermediate node, P6 m2 、P6 m4 For the 6th layer intermediate node, P4 out 、P5 out It is the feature output of layer 4 and layer 5. Similar to layer 6, the feature output of layer 5 enters the adaptive dual-squeeze excitation channel attention mechanism.

[0086] In this embodiment, the Adaptive DualSqueeze Excitation Channel Attention Module (ADSE) in step S3 is as follows: Figure 5 As shown in the figure; use global average pooling and global standard deviation pooling to squeeze the feature information into channels and obtain the statistical information of each channel; then use a fully connected layer and activation function ReLU to perform channel compression, and a fully connected layer to restore the number of channels to obtain the relevant information between channels; then use the H-Sigmoid activation function for nonlinear normalization to obtain the attention weight of each channel; finally, multiply the attention weight of each channel with the feature map, and add the self-learning weight factor to balance the information loss caused by attention, and add it to the original feature map.

[0087] ADSE performs weighted processing on the feature map according to the learned channel attention weights, giving lower weights to feature information with lower relevance to the target, and vice versa, giving higher weights to feature information with lower relevance to the target, highlighting important target information in the features and suppressing other irrelevant information such as background, thereby improving the detection capability of this technical method for weak infrared target detection on water.

[0088] In this embodiment, the specific process of ADSE in step S3 includes:

[0089] (1) Order is the input feature, the size is C×H×W; the feature is subjected to global average pooling and global standard deviation pooling respectively, and we can get

[0090]

[0091]

[0092] Then, the above features are spliced ​​together to obtain the features t c =Cat([μ c ,σ c ]), where Cat(·) represents the channel concatenation operation on all feature maps.

[0093] (2) Use the excitation operation to reduce the amount of computation and obtain the channel attention map; that is, perform a full connection with parameter W0 on the feature T to obtain a feature with dimension C / 2×1×1; then use the ReLU activation function δ and restore it to the original channel dimension C×1×1 of the image input through a fully connected layer with parameter W1; finally, use the H-Sigmoid activation function for normalization to obtain the following attention weights S for each channel:

[0094] S=σ(W1δ(W0T))

[0095] (3) Multiply the attention weight S of each channel with the feature map, and add the self-learned weight factor α to the original feature map to obtain the following output features

[0096] E=α(SX)+X

[0097] The weight factor α is a learnable parameter; its initial value can be set to 0, and a larger and more appropriate weight can be obtained during network training.

[0098] S4: The multi-scale fusion features output by the dense bidirectional pyramid network are input into the classification prediction network and the prediction box regression network respectively to obtain the position and category information of all targets in the water infrared image.

[0099] In specific implementation, the classification prediction network and prediction box regression network described in step S4 use the fusion features output by the dense bidirectional feature pyramid network to respectively predict the category and bounding box of the water infrared target, and the two networks share the same feature network weights; then the predicted target results are sorted by score and non-maximum suppression is performed to filter out the boxes and scores of each category with scores greater than the confidence threshold; non-maximum suppression is performed according to the position and score of the box, and finally the position and category information of all targets in the water infrared image are obtained.

[0100] In the above embodiment, S2 extracts multi-scale deep features of the target from the input image to be detected, S3 fuses the multi-scale feature information, and S4 classifies and locates the target in the underwater infrared image, outputting the final detected target. In the actual underwater infrared target detection process, the sequence of S1, S2, S3, and S4 can be repeated to achieve target detection. Furthermore, this embodiment requires pre-training the network model to better complete the target detection process.

[0101] For network training, the SMD dataset, VAIS dataset and LSOTB-TIR dataset are used, and the data of the water infrared scene in these three datasets are selected; the dataset is randomly divided into training set, validation set and test set, accounting for 80%, 10% and 10% of the data respectively.

[0102] Due to the limited number of overwater infrared datasets, this method uses data augmentation methods such as random scaling, random cropping, and random placement to enrich the overwater infrared dataset and improve infrared target detection. Before network training, this method automatically sets the initial anchor box size based on the target detection dataset.

[0103] For network training, we set the batch size to 16, the number of epochs to 1000, the initial learning rate to 0.001, the re-decay rate to 0.0005, the number of iterations to 200, and the Adam optimizer. We first froze the feature extraction network front end and trained it for 100 epochs, then unfrozen and trained the entire network for 150 epochs. Finally, we trained and adjusted the parameters of the entire network.

[0104] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0105] The present 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 the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting infrared targets on water based on a dense bidirectional pyramid network, characterized in that: The following steps are involved: S1, adaptively scaling the collected water infrared images to a uniform size; S2. Input the image into the feature extraction network to extract multi-scale depth feature maps of various levels; the feature extraction network is an improved EfficientNetV2 network, including a Fused-MBConv module and an MBConv module. The Fused-MBConv module is a residual network structure. The residual part consists of a 3×3 ordinary convolution, a 3×3 depth-separable convolution, a residual structure of the SE channel attention mechanism, a 1×1 ordinary convolution and a Dropout layer; the SE channel attention mechanism includes a global average pooling, two fully connected layers and a Sigmoid activation function; the improved EfficientNetV2 network includes 7 stages. The first stage is a 3×3 ordinary convolution with a step size of 2, including BN and SiLu, and the output is Layer; the second to fifth stages are repeated stacking of Fused-MBConv modules. The second stage repeats the Fused-MBConv module 2 times with a step size of 1, and the third stage repeats the Fused-MBConv module 4 times with a step size of 2. The output is layer, the fourth stage repeats the Fused-MBConv module 4 times with a step size of 2, and the output is layer, the fifth stage repeats the Fused-MBConv module 6 times with a step size of 2, and the output is layer; in the sixth stage, the MBConv module is repeated 9 times with a step size of 2, and the output is layer; the seventh stage consists of a normal 1×1 convolutional layer, an average pooling layer and a fully connected layer; Layer The input of the dense bidirectional pyramid network is the feature extraction network extracted by layer, layer, layer, Layer and Layer feature information; S3. Input the obtained multi-scale feature maps of various levels into a dense bidirectional pyramid network to fuse the multi-scale features. The dense bidirectional pyramid network adds an adaptive dual-squeeze excitation channel attention mechanism before the multi-scale fusion feature output; the fusion of the multi-scale features adopts a multi-scale feature fusion method of top-down, bottom-up and jump dense connection, and the dense bidirectional feature pyramid network adds different normalized weights to each node input feature; the jump dense connection adopts an interval jump connection method, The node receives the data from the previous node with a depth of The sum of the characteristic information of the distance, where , , ; The adaptive dual-squeeze excitation channel attention mechanism includes: Use global average pooling and global standard deviation pooling to squeeze the feature information into channels and obtain statistical information of each channel; use a fully connected layer and the activation function ReLU to perform channel compression, and a fully connected layer to restore the number of channels to obtain the relevant information between channels; use the H-Sigmoid activation function for nonlinear normalization to obtain the attention weight of each channel; multiply the attention weight of each channel by the feature map, add a self-learned weight factor to balance the information loss caused by attention, and add it to the original feature map; S4. The multi-scale fusion features output by the dense bidirectional pyramid network are input into the classification prediction network and the prediction box regression network respectively to obtain the position and category information of all targets in the water infrared image.

2. The method for detecting infrared targets on water based on a dense bidirectional pyramid network according to claim 1, wherein: The normalized weight method is described as follows: in, are the input features, is the output fusion feature, and is a learnable weight, which is limited to between 0 and 1, is the learning rate, with a value of 0.001, and is the number of network layers.

3. The method for detecting infrared targets on water based on a dense bidirectional pyramid network according to claim 1, wherein: The specific process of the dual-squeeze excitation channel attention mechanism includes: make is the input feature, the size is , perform global average pooling and global standard deviation pooling on the feature respectively, and get Perform channel splicing on the above features to obtain features , ,in, Indicates channel concatenation operation on all feature maps; Use the excitation operation to reduce the amount of computation and obtain the channel attention map, including parameterizing the feature T as The full connection of characteristics; Use ReLU activation function , and pass the parameter as The fully connected layer restores the original channel dimension of the image input ; After normalization using the H-Sigmoid activation function, the attention weights of each channel are obtained as follows: : The attention weight of each channel Multiply it with the feature map and add the self-learning weight factor Then add it to the original feature map to get the following output features : Among them, the weight factor is a learnable parameter.

4. The method for detecting infrared targets on water based on a dense bidirectional pyramid network according to claim 1, wherein: In step S4, the classification prediction network and the prediction box regression network use the fusion features output by the dense bidirectional feature pyramid network to predict the category and bounding box of the water infrared target respectively, and the two networks share the same feature network weights; Sorting and non-maximum suppression are performed on the predicted target results to filter out the boxes and scores of each category whose scores are greater than the confidence threshold; Non-maximum suppression is performed based on the position and score of the frame to obtain the position and category information of all targets in the water infrared image.

Images