Infrared iceberg target detection method based on adaptive feature enhancement interaction network

Abstract

The application relates to an infrared iceberg target detection method based on an adaptive feature enhancement interaction network and relates to the field of iceberg detection. The application is aimed at solving the problem of low detection accuracy of existing target detection methods for infrared iceberg target detection. The application comprises the following steps: acquiring an infrared iceberg image to be identified, inputting the infrared iceberg image to be identified into a trained ADENet network, acquiring an infrared iceberg feature map set E={E1, E2, E3, E4, E5} composed of infrared iceberg feature maps of different resolutions by an ADENet network encoder, extracting target global features by using the feature map E5 in a global contrast enhancement module, obtaining a feature map and guide features E guide ; extracting target semantic features by using the feature E guide in a gating semantic feature extraction module, obtaining a feature map j=2, 3, 4; extracting target shallow layer detail features by using the feature map E1 in a detail feature extraction module, obtaining a feature map, and using a decoder for decoding the feature map and aggregating to obtain a target recognition result in the infrared iceberg image. The application is used for detecting infrared iceberg targets.

Description

Technical Field

[0001] This invention relates to the field of iceberg detection, and in particular to an infrared iceberg target detection method based on an adaptive feature enhancement interactive network. Background Technology

[0002] In recent years, the melting rate of Arctic sea ice has accelerated, and activities in the Arctic region have gradually increased. The Arctic shipping routes, with their advantages in economic and time costs, are playing an increasingly prominent role in international trade and shipping development. However, the Arctic Ocean is home to a large number of icebergs, with new icebergs constantly breaking off from glaciers each year. These include small icebergs that are easily overlooked by detection systems and crew members. These icebergs move irregularly with the waves on the turbulent sea surface, posing a significant threat to Arctic shipping routes. Failure to detect them in a timely manner at a safe distance can endanger ships. Therefore, strengthening the observation of icebergs around ships is of great importance to ensuring safe navigation.

[0003] Currently, various tools have been developed for iceberg detection. Model prediction or satellite methods are often used for monitoring large icebergs (over 100 meters wide). All-weather remote sensing methods, such as spaceborne polarimetric synthetic aperture radar (SAR) and airborne radar, are limited by resolution and revisit time and are mainly used to detect small to medium-sized icebergs (greater than 10 meters wide). These iceberg targets are relatively easy to detect due to their significant size, while the more common, relatively small icebergs (namely growlers and bergy bits) are the main threat to ships. For these targets, the crew generally makes judgments based on shipboard detection systems combined with experience. Detection systems include marine radar, optical, and infrared sensors. However, due to the extremely harsh sea conditions in the Arctic, marine radar signals are often submerged in sea clutter in radar images, making it difficult to distinguish them from clutter. Optical sensors are also limited by fog and low-light conditions, which greatly reduces detection effectiveness. Therefore, in harsh marine conditions, infrared iceberg detection is an important supplementary method for iceberg monitoring in the Arctic Ocean.

[0004] Infrared sensors can capture the contrast differences between icebergs and the surrounding seawater background. They are highly immune to fog, haze, and other weather conditions, enabling all-day operation and working in low-light conditions. Infrared detection technology has proven to be an effective tool for iceberg detection. However, infrared iceberg target detection faces many challenges. First, as the detection distance increases, the imaging contrast between the iceberg and the seawater decreases, and low-contrast infrared iceberg data increases the difficulty of detection. Additionally, some strong waves and wakes have higher grayscale and local contrast than the target, making them easily misdetected as real targets. Furthermore, iceberg targets have varied and irregular shapes, with limited pixel and texture information, resulting in limited usable features. Currently, many traditional methods have been proposed for infrared small target detection, including filter-based methods, methods based on target and background feature calculations, and optimization methods using low-rank sparse matrix operations. However, these traditional methods often rely on manually extracted image features and pre-set hyperparameters, and they only perform well in simple scenarios. Therefore, deep learning methods have been increasingly applied to infrared detection tasks, mainly including target detection methods based on deep adversarial learning networks and methods that modularize LCM methods into networks for target detection. However, these methods are prone to losing features when detecting small targets, resulting in undetected targets. Therefore, a method combining the semantic features of small targets with their edge information has been proposed for target detection. This method overcomes the shortcoming of deep learning-based target detection methods that easily lose features of small targets. However, infrared iceberg images are complex scenes with low contrast and severe sea wave clutter, such as… Figure 1 As shown, existing small target detection methods cannot adapt to dim targets and strong sea clutter scenarios, resulting in false alarms and false detections, thus leading to low accuracy in infrared iceberg target detection. Summary of the Invention

[0005] The purpose of this invention is to address the problem of low accuracy in infrared iceberg target detection using existing target detection methods, and to propose an infrared iceberg target detection method based on an adaptive feature enhancement interactive network.

[0006] The infrared iceberg target detection method based on adaptive feature enhancement interactive network is as follows:

[0007] Acquire an infrared image of the iceberg to be identified, input the image into the trained ADENet network, and obtain the target recognition result;

[0008] The training set of the ADENet network consists of the original infrared iceberg image and the infrared iceberg image after the target location is labeled.

[0009] The ADENet network includes: an encoder, a global contrast enhancement module, a gated semantic feature extraction module, a detail feature extraction module, and a decoder;

[0010] The encoder is a VGG network, and the encoder is used to acquire an infrared iceberg feature map set E = {E1, E2, E3, E4, E5} consisting of five infrared iceberg image feature maps of different resolutions.

[0011] in, i∈{1,2,3,4,5},E i C is the feature map of the i-th infrared iceberg image. i It is E i The number of channels, H i It is E i Height, W i It is E i The width;

[0012] The global contrast enhancement module uses feature map E5 to extract global features of the target and obtain the feature map. and guiding feature E guide ;

[0013] The gated semantic feature extraction module is used for target semantic features, utilizing feature map E4 and feature E guide Extracting and obtaining feature maps Using feature maps E3 and feature map to obtain feature map Using feature maps Feature map obtained from feature map E2

[0014]

[0015] The detailed feature extraction module utilizes feature map E1, Extract shallow detail features of the target to obtain a feature map.

[0016] The decoder is used to process the feature map. Decode and aggregate to obtain target recognition results in infrared iceberg images.

[0017] Furthermore, the feature map Obtained through the following methods:

[0018]

[0019] Conv() is a 3×3 convolution operation. It is an intermediate feature map. Conv 1×1 () represents a 1×1 convolution operation. 3×3 () represents a 3×3 convolution operation. 5×5 () represents a 5×5 convolution operation. 7×7() is a 7×7 convolution operation, and Concate() is a feature connection operation.

[0020] Furthermore, the guiding feature E guide Specifically:

[0021]

[0022] SA() is a spatial attention operation.

[0023] Furthermore, the gated semantic feature extraction module is used to extract target semantic features using feature map E4 and feature E guide Extracting and obtaining feature maps Using feature maps E3 and feature map to obtain feature map Using feature maps Feature map obtained from feature map E2 Specifically:

[0024]

[0025] Here, up() is the upsampling operation. It is an intermediate feature map. It is a point-by-point multiplication operation of spatial dimensions, where j takes the values ​​2, 3, or 4.

[0026] Furthermore, Specifically:

[0027]

[0028]

[0029] Among them, Conv di1 This is a dilated convolution operation with a dilated rate of 1. `stack()` is a feature map stacking operation, and `mean()` is a feature map mean calculation operation. It is an intermediate feature map, Conv di3 () is a dilated convolution operation with a dilated rate of 3. di5 () is a dilated convolution operation with a dilated rate of 5. di7 () is a dilated convolution operation with a dilated rate of 7, x gatek It is the characteristic of the output of the kth threshold mechanism operation, where k takes the values ​​1, 3, or 5.

[0030] Furthermore, the characteristics of the output of the k-th threshold mechanism operation are specifically as follows:

[0031]

[0032] Here, gate() is the threshold mechanism operation, and Sigmoid() is the Sigmoid function.

[0033] Furthermore, the detailed feature extraction module utilizes feature map E1, Extract shallow detail features of the target to obtain a feature map. Specifically:

[0034]

[0035] in, These are intermediate feature maps. Aver() is the average pooling operation, Max() is the max pooling operation, and CA() is the channel attention operation. is the feature map output by the gated semantic feature extraction module, and ⊙ is the pointwise multiplication operation of the channel dimension.

[0036] Furthermore, the decoder is used for feature maps The target recognition results in the infrared iceberg image are obtained by decoding and aggregating the data. Specifically:

[0037] A1. Using feature maps Obtain feature map D5:

[0038]

[0039] A2. Initialize i = 5;

[0040] A3. Utilizing feature maps and D i Obtain feature map D i-1 ;

[0041] A4. Determine if i is 2. If i = 2, output the target recognition result D1 in the infrared iceberg image. If i > 2, set i = i-1 and return to A3.

[0042] Furthermore, the feature map in A3... and D i Obtain feature map D i-1 Specifically:

[0043]

[0044] Here, BN() is the batch normalization operation, and ReLU() is the ReLU activation function.

[0045] Furthermore, the loss function of the ADENet network is a hybrid loss function. mix Specifically:

[0046]

[0047] Among them, l BCE (D i ,GT) is the binary cross-entropy loss function, l IoU (D i ,GT) is the IoU loss function, and GT is the ground truth value of the target in the infrared iceberg image.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention proposes an infrared iceberg target detection method based on an adaptive feature enhancement interactive network (ADENet). The method utilizes a global contrast enhancement module to adaptively enhance low-contrast images centered on the target, recovering latent global features. The feature guidance branch within this module provides guidance for the extraction of target semantic features. Furthermore, the invention designs a detail feature extraction module to adaptively enhance detailed features such as target edges, adaptively extracting target edge contours and enhancing image detail features. This approach avoids false alarms and false detections caused by dim targets and strong sea clutter in infrared iceberg images, thus improving the accuracy of infrared iceberg target detection. Attached Figure Description

[0050] Figure 1 A three-dimensional view of infrared iceberg targets and ocean clutter;

[0051] Figure 2 Interactive network structure diagram for adaptive feature enhancement;

[0052] Figure 3 Here is a structural diagram of the gated semantic feature extraction module;

[0053] Figure 4 This is a structural diagram of the global contrast enhancement module;

[0054] Figure 5 This is a schematic diagram illustrating the computational principle of the detail feature extraction module.

[0055] Figure 6(a) shows a comparison of the detection performance of ADENet with other methods on the SIRST dataset;

[0056] Figure 6(b) shows a comparison of the detection performance of ADENet with other methods on the SFIIT dataset. Detailed Implementation

[0057] Specific implementation method one: as follows Figure 2 As shown, the specific process of the infrared iceberg target detection method based on the adaptive feature enhancement interactive network in this embodiment is as follows: acquire the infrared iceberg image to be identified, input the infrared iceberg image to be identified into the trained ADENet network, and obtain the target recognition result.

[0058] The training set of the ADENet network consists of the original infrared iceberg image and the infrared iceberg image after the target location is labeled.

[0059] The ADENet network includes: an encoder, a global contrast enhancement module, a gated semantic feature extraction module, a detail feature extraction module, and a decoder;

[0060] The encoder is a VGG network used to downsample the input infrared iceberg image and map the image features to a high-dimensional space to obtain an infrared iceberg feature map set E = {E1, E2, E3, E4, E5} consisting of feature maps of five infrared iceberg images with different resolutions; the input of the encoder is an infrared iceberg image with a size of 512×512×3.

[0061] in, i∈{1,2,3,4,5},E i This is the feature map of the i-th infrared iceberg image, C1=64, C2=128, C3=256, C4=512, C5=512, H1=W1=512, H2=W2=256, H3=W3=128, H4=W4=64, H5=W5=32, C i It is E i The number of channels, H i It is E i Height, W i It is E i The width;

[0062] The Global Contrast Enhancement (GCE) module extracts global features of the target using feature map E5 to obtain the feature map. and guiding feature E guide ;

[0063] The gated semantic feature extraction module is used for target semantic features, utilizing feature map E4 and feature E guide Extracting and obtaining feature maps Using feature maps E3 and feature map to obtain feature map Using feature maps Feature map obtained from feature map E2

[0064]

[0065] The detail feature extraction module DFE utilizes feature map E1, Extract shallow detail features of the target to obtain a feature map.

[0066] The decoder is used to process the feature map. Decode and aggregate to obtain target recognition results in infrared iceberg images.

[0067] Specific implementation method two: such as Figure 4 As shown, the global contrast enhancement module includes: a contrast enhancement branch and a feature guidance branch;

[0068] The contrast enhancement branch is used to acquire feature maps. The contrast enhancement branch sequentially includes: a 3×3 first convolutional layer, a 1×1 second convolutional layer, a 3×3 third convolutional layer, a 5×5 fourth convolutional layer, a 7×7 fifth convolutional layer, a first feature connection layer, a 3×3 sixth convolutional layer, and a feature map. Output layer;

[0069] The second, third, fourth, and fifth convolutional layers are parallel convolutional layers, and the input feature maps are all the output feature maps of the first convolutional layer;

[0070] The first convolutional layer of the 3×3 structure is obtained using feature map E5. Specifically:

[0071]

[0072] The inputs to the 1×1 first convolutional layer, the 3×3 second convolutional layer, the 5×5 third convolutional layer, and the 7×7 fourth convolutional layer are all feature maps. The output feature map is passed through a feature map connection layer to obtain a new feature map. Specifically:

[0073]

[0074] Concate() is a feature concatenation operation;

[0075] The input to the sixth convolutional layer of the 3×3 matrix is ​​the feature map. After convolution, the input is fed into the feature map. Output layer;

[0076] The feature map The output layer will pass through the convolutional feature maps and Summing elements to obtain the feature map Specifically:

[0077]

[0078] Conv() is a 3×3 convolution operation;

[0079] The feature-guided branch utilizes feature maps. and feature map Obtain guiding feature E guide ;

[0080] The feature-guided branch includes: a first feature connection layer, a 3×3 first convolutional layer, a spatial attention mechanism layer, and a 3×3 second convolutional layer;

[0081] The feature-guided branch is processed as follows:

[0082]

[0083] Among them, E guide It is a guiding feature, and SA is a spatial attention operation.

[0084] In this implementation, E5 has the lowest resolution and lacks shallow-level detail features, but it contains global contrast features and large-scale contours, among other global information. The global contrast enhancement module is divided into a contrast enhancement branch and a feature guidance branch. First, E5 is input into the contrast enhancement branch, where information is obtained through parallel convolutions with kernels of 1×1, 3×3, 5×5, and 7×7. Then, the feature maps from the parallel convolutions are concatenated. Subsequently, the input and the output of the contrast enhancement branch are input into the feature guidance branch, which emphasizes the image contrast features through a spatial attention mechanism, and outputs the guiding feature E. guide This provides guidance information for the gated semantic feature extraction module. In this embodiment, the first feature connection layer, the first convolutional layer, and the second convolutional layer of the macro are not the same convolutional layer, but rather the first feature connection layer, the first convolutional layer, and the second convolutional layer under the corresponding branch.

[0085] Specific implementation method three: such as Figure 3 As shown, the gated semantic feature extraction module includes: feature map Acquisition of units and feature maps Acquisition of units and feature maps Acquisition of units and feature maps Acquisition unit, feature map stacking unit, feature map averaging unit, Feature map output unit;

[0086] The feature map The acquisition unit includes, in sequence: a first dilated convolution subunit with a dilated rate of 1 and a first feature summation subunit;

[0087] Feature map E j Inputting the first dilated convolutional subunit with a dilated rate of 1, the E after dilation and convolution... j With feature E j The feature map is obtained after the first feature summation subunit. Specifically:

[0088]

[0089] Among them, Convdi1 This is the first dilated convolution operation with a dilated rate of 1;

[0090] The feature map The acquisition unit sequentially includes: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolutional layer with a dilated rate of 3, and a second feature summation subunit;

[0091] The threshold mechanism sub-unit includes, in sequence: a 3x3 convolutional sub-layer, a sigmoid function sub-layer, and a spatial dimension pointwise multiplication sub-layer; the input of the 3x3 convolutional sub-layer is the feature map output from the previous layer. The output of the 3x3 convolutional sublayer serves as the input to the sigmoid function sublayer. The output of the sigmoid function sublayer, combined with the output of the previous feature summation layer, becomes the input to the spatial dimension pointwise multiplication sublayer. The output of the spatial dimension pointwise multiplication sublayer is the output of the threshold mechanism subunit. Specifically:

[0092]

[0093] Where k takes the values ​​1, 3, or 5. Pick x gatek It is the output characteristic of the threshold mechanism subunit, and Sigmoid represents the Sigmoid function. For point-by-point multiplication operations in spatial dimensions, gate() is a threshold mechanism operation;

[0094] Feature map The first threshold mechanism subunit in the acquisition unit takes the feature map as its input. The output is x gate1 The output x of the first threshold mechanism subunit gate1 With feature map E j The first feature summation subunit obtains the summation result, which is then passed through a first dilated convolutional layer with a dilation rate of 3. The output of the first dilated convolutional layer with a dilation rate of 3, along with the output of the second feature summation subunit, is then passed through another feature summation subunit to obtain the feature map. Specifically:

[0095]

[0096] Among them, Conv di3 () is a dilated convolution operation with a dilated rate of 3;

[0097] Feature map The acquisition unit sequentially includes: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolutional layer with a dilated rate of 5, and a second feature summation subunit; feature map Acquisition layer processing flow and feature map The processing flow for the acquisition layer is similar, specifically as follows:

[0098]

[0099] Among them, Conv di5 () is a dilated convolution operation with a dilated rate of 5;

[0100] Feature map Each acquisition unit sequentially includes: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolutional layer with a dilated rate of 7, and a second feature summation subunit; feature map Acquisition layer processing flow and feature map The processing flow for the acquisition layer is similar, specifically as follows:

[0101]

[0102] Among them, Conv di7 () is a dilated convolution operation with a dilated rate of 7;

[0103] The feature map stacking unit is used for stacking feature maps. Perform a stacking operation to obtain the feature map stacking result;

[0104] The feature map averaging unit is used to perform a feature map averaging operation on the feature map stacking result to obtain the feature map. Specifically:

[0105]

[0106] Among them, stack() is the feature map stacking operation, and mean() is the operation to calculate the average value of the feature maps;

[0107] The The feature map output unit is used to process the feature map. Perform channel attention and spatial attention operations to obtain feature maps. Then use feature maps Feature map E j Feature map Obtain feature maps Specifically:

[0108]

[0109]

[0110] Here, up() is the upsampling operation.

[0111] In this embodiment, multi-scale information of the target is first obtained through dilated convolution operations at different dilation sizes. A threshold mechanism is introduced to add an information propagation channel between adjacent dilated convolution branches. This mechanism incorporates feature maps generated by convolutional layers and Sigmoid function filtering. Feature maps in this embodiment... Acquisition of units and feature maps Acquisition of units and feature maps Acquisition of units and feature maps The first dilated convolution subunit, the first feature summation subunit, the second feature summation subunit, and the first threshold mechanism subunit in the acquisition unit do not refer to the same subunit, but rather to which subunit in their respective corresponding layers.

[0112] Specific implementation method four: such as Figure 5 As shown, the detailed feature extraction module includes: an element separation unit and a feature map. Acquisition of units and feature maps Acquisition of units and feature maps Acquisition of units and feature maps Acquisition of units and feature maps Acquisition Unit;

[0113] The element separation unit includes a first 3×3 convolutional layer;

[0114] Feature map E1 input is obtained by the first 3×3 convolutional layer. Specifically:

[0115]

[0116] The feature map The acquisition unit sequentially includes: an upsampling layer, a spatial dimension pointwise multiplication operation layer, a spatial attention layer, and a feature map summation layer; the feature map The input to the acquisition unit is the feature map. and feature map First, after passing through an upsampling layer, it has the same resolution as E1. Then, after upsampling... and The input spatial dimension pointwise multiplication operation layer, the output of the spatial dimension pointwise multiplication operation layer is the input of the spatial attention layer, and the output of the spatial attention layer is... Input feature map summation layer to obtain feature map Specifically:

[0117]

[0118] The feature map The acquisition unit includes: feature map Obtain sub-units and feature maps Obtain sub-units;

[0119] The feature map The sub-units are sequentially comprised of: a channel attention layer, a layer of element-wise multiplication of channel dimensions, and a feature map summation layer;

[0120] The input to the channel attention layer is The process sequentially passes through a channel attention layer, a layer of element-wise multiplication along the channel dimensions, and a feature map summation layer to obtain the feature map. Specifically:

[0121]

[0122] Where ⊙ is the pointwise multiplication operator for channel dimensions, and CA() is the channel attention operation;

[0123] The feature map The acquisition subunit sequentially includes: a spatial attention layer, a spatial dimension pointwise multiplication operation layer, and a feature map summation layer; the feature map Obtain the sub-unit input as a feature map Feature maps are obtained sequentially through a spatial attention layer, a pointwise multiplication layer of spatial dimensions, and a feature map summation layer. Specifically:

[0124]

[0125] The feature map The acquisition unit includes: feature map Obtain sub-units and feature maps Obtain sub-units;

[0126] The feature map The acquisition sub-unit includes: a channel attention layer, a channel-dimensional pointwise multiplication layer, and a feature map summation layer; the feature map... The input to the subunit is the feature map. The feature map is obtained by sequentially passing through the channel attention layer, the pointwise multiplication layer of the channel dimension, and the feature map summation layer. Specifically:

[0127]

[0128] The feature map The acquisition sub-unit sequentially includes: a spatial attention layer, a spatial dimension pointwise multiplication operation layer, and a feature map summation layer; the feature map Obtain the sub-unit input as a feature map Feature maps are obtained sequentially through a spatial attention layer, a pointwise multiplication layer of spatial dimensions, and a feature map summation layer. Specifically:

[0129]

[0130] The feature map The acquisition unit includes: average pooling layer, max pooling layer, and feature map subtraction operation layer;

[0131] The average pooling layer is used to process the feature map. Perform average pooling operation;

[0132] The max pooling layer is used to process the feature map. Perform average pooling operation;

[0133] The feature map subtraction layer performs a subtraction operation between the output of the average pooling layer and the output of the max pooling layer to obtain the feature map. Specifically:

[0134]

[0135] Among them, Aver() is the average pooling operation, and Max() is the max pooling operation;

[0136] The feature map The acquisition unit includes: a feature connection layer, a sigmoid activation function layer, and a feature map summation layer;

[0137] The input to the feature map connection layer is a feature map. The feature map is obtained by passing the data through a Sigmoid activation function layer and a feature map summation layer in sequence. Specifically:

[0138]

[0139] Concate() is a feature map connection operation.

[0140] In this step, average pooling and max pooling are applied respectively, followed by difference processing to adaptively perceive shallow details such as edges, without requiring separate supervision of edge ground plane. Then, skip connections are used to combine the differenced feature map with the two feature maps before and after the difference to form the output of the detail feature extraction module. Finally, we extract the differenced feature map... and the feature map of the two branches The features are connected and adaptively aggregated using 1×1 convolution. Finally, they are added to the input shallow feature map E1 to obtain the output of the detail feature extraction module DFE.

[0141] Feature map in this embodiment Obtain sub-units and feature maps Obtain sub-units and feature maps Obtain sub-units and feature maps The channel attention layer, channel dimension pointwise multiplication element layer, feature map summation layer, spatial attention layer, and spatial dimension pointwise multiplication operation layer mentioned in the sub-unit are not the same layer, but rather the channel attention layer, channel dimension pointwise multiplication element layer, feature map summation layer, spatial attention layer, and spatial dimension pointwise multiplication operation layer in their respective sub-units.

[0142] Specific Implementation Method Five: The decoder is used to process the feature map The target recognition results in the infrared iceberg image are obtained by decoding and aggregating the data. Specifically:

[0143] A1. Using feature maps Obtain feature map D5:

[0144]

[0145] A2. Initialize i = 5;

[0146] A3. Utilizing feature maps and D i Obtain feature map D i-1 :

[0147]

[0148] Where BN() is the batch normalization operation and ReLU() is the ReLU activation function;

[0149] A4. Determine if i is 2. If i = 2, output the target recognition result D1 in the infrared iceberg image. If i > 2, set i = i-1 and return to A3.

[0150] Specific Implementation Method Six: The ADENet network training employs the following hybrid loss function. mix :

[0151]

[0152] Among them, l BCE (D i ,GT) is the binary cross-entropy loss function, l IoU (D i ,GT) is the IoU loss function, and GT is the ground truth value of the target in the infrared iceberg image.

[0153] Example: To verify the beneficial effects of the present invention, the ADENet network proposed in this invention was compared with existing networks. The results are shown in Table 1, and the effect comparison diagram is shown below. Figures 6(a)-6(b) As shown;

[0154] Table 1

[0155]

[0156] As shown in Figure 6(a), on the publicly available infrared small target dataset SIRST, ADENet achieves the best evaluation metrics in nIoU, Fa, and Prec, demonstrating better small target segmentation capabilities than other methods. Furthermore, due to the contrast enhancement module in the ADENet network, the ADENet method reduces false alarms in target detection, resulting in a higher F1 score. As shown in Figure 6(b), on the infrared iceberg target dataset SFIIT, the ADENet method validates its superiority in iceberg target detection across various data points. Its IoU and nIoU metrics are significantly higher than other methods, and it achieves a balance in Prec and Rec metrics, enabling it to detect iceberg targets with good control over false alarms and obtaining the best F1 score.

Claims

1. An infrared iceberg target detection method based on an adaptive feature enhancement interactive network, characterized in that... The specific process of the method is as follows: acquire an infrared iceberg image to be identified, input the infrared iceberg image to be identified into a trained ADENet network, and obtain the target recognition result; The training set of the ADENet network consists of the original infrared iceberg image and the infrared iceberg image after the target location is labeled. The ADENet network includes: an encoder, a global contrast enhancement module, a gated semantic feature extraction module, a detail feature extraction module, and a decoder; The encoder is a VGG network, and the encoder is used to acquire an infrared iceberg feature map set E = {E1, E2, E3, E4, E5} consisting of five infrared iceberg image feature maps of different resolutions. in, i∈{1,2,3,4,5},E i C is the feature map of the i-th infrared iceberg image. i It is E i The number of channels, H i It is E i Height, W i It is E i The width; The global contrast enhancement module uses feature map E5 to extract global features of the target and obtain the feature map. and guiding feature E guide ; The gated semantic feature extraction module is used for target semantic features, utilizing feature map E4 and feature E guide Extracting and obtaining feature maps Using feature maps E3 and feature map to obtain feature map Using feature maps Feature map obtained from feature map E2 The detailed feature extraction module utilizes feature map E1, Extract shallow detail features of the target to obtain a feature map. The decoder is used to process the feature map. Decode and aggregate to obtain target recognition results in infrared iceberg images.

2. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 1, characterized in that: The feature map Obtained through the following methods: Conv() is a 3×3 convolution operation. It is an intermediate feature map, Conv 1×1 () represents a 1×1 convolution operation. 3×3 () represents a 3×3 convolution operation. 5×5 () represents a 5×5 convolution operation. 7×7 () is a 7×7 convolution operation, and Concate() is a feature connection operation.

3. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 2, characterized in that: The guiding feature E guide Specifically: SA() is a spatial attention operation.

4. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 3, characterized in that: The gated semantic feature extraction module is used to extract target semantic features using feature map E4 and feature E guide Extracting and obtaining feature maps Using feature maps E3 and feature map to obtain feature map Using feature maps Feature map obtained from feature map E2 Specifically: Here, up() is the upsampling operation. It is an intermediate feature map. It is a point-by-point multiplication operation of spatial dimensions, where j takes the values ​​2, 3, or 4.

5. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 4, characterized in that: Specifically: Among them, Conv di1 This is a dilated convolution operation with a dilated rate of 1. `stack()` is a feature map stacking operation, and `mean()` is a feature map mean calculation operation. It is an intermediate feature map, Conv di3 () is a dilated convolution operation with DilatedRate = 3. di5 () is a dilated convolution operation with a dilated rate of 5. di7 () is a dilated convolution operation with a dilated rate of 7, x gatek It is the characteristic of the output of the kth threshold mechanism operation, where k takes the values ​​1, 3, or 5.

6. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 5, characterized in that: The characteristics of the output of the k-th threshold mechanism operation are as follows: Here, gate() is the threshold mechanism operation, and Sigmoid() is the Sigmoid function.

7. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 6, characterized in that: The detailed feature extraction module utilizes feature map E1, Extract shallow detail features of the target to obtain a feature map. Specifically: in, These are intermediate feature maps. Aver() is the average pooling operation, Max() is the max pooling operation, and CA() is the channel attention operation. is the feature map output by the gated semantic feature extraction module, and ⊙ is the pointwise multiplication operation of the channel dimension.

8. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 7, characterized in that: The decoder is used to process the feature map. The target recognition results in the infrared iceberg image are obtained by decoding and aggregating the data. Specifically: A1. Using feature maps Obtain feature map D5: Where BN() is the batch normalization operation and ReLU() is the ReLU activation function; A2. Initialize feature icon number i = 5; A3. Utilizing feature maps and D i Obtain feature map D i-1 ; A4. Determine if i is 2. If i = 2, output the target recognition result D1 in the infrared iceberg image. If i > 2, set i = i-1 and return to A3.

9. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 8, characterized in that: The use of feature maps in A3 and D i Obtain feature map D i-1 Specifically: ReLU() is the ReLU activation function.

10. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 9, characterized in that: The loss function of the ADENet network is a hybrid loss function. mix Specifically: Among them, l BCE (D i ,GT) is the binary cross-entropy loss function, l IoU (D i ,GT) is the IoU loss function, and GT is the ground truth value of the target in the infrared iceberg image.

Images