Infrared iceberg target detection method based on adaptive feature enhanced interactive network
Patent Information
- Application Number
- CN202411091641.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-09
AI Technical Summary
The existing infrared iceberg target detection methods are difficult to accurately detect small icebergs under harsh marine conditions, which are prone to false alarms and false detection, resulting in low detection accuracy.
Using an adaptive feature enhancement interactive network (ADENet)-based method, the global contrast enhancement module and the detail feature extraction module are used to adaptively enhance low-contrast images, recover potential global features and detailed features, and improve the accuracy of object detection.
It effectively improves the accuracy of infrared iceberg target detection, reduces false alarms and false detection, and adapts to iceberg target detection in complex scenarios.
Smart Images

Figure CN119006789A_ABST
Abstract
Description
Technical Field
[0001] The 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 Art
[0002] In recent years, the melting speed of Arctic sea ice has accelerated, and activities in the Arctic region have gradually increased. The Arctic route has become increasingly prominent in the development of international trade and shipping due to its superiority in economic cost and time cost. There are a large number of icebergs floating on the Arctic sea surface, and new icebergs are constantly breaking off from the glaciers every year. These include small icebergs that are easily ignored by detection systems and crew members. They move irregularly with the waves on the turbulent sea surface, which is a huge threat to the Arctic route. If they cannot be discovered in time at a safe distance, they will bring risks to the ships on the road. Therefore, strengthening the observation of icebergs around ships is of great significance to ensure the safe navigation of ships.
[0003] Various tools for iceberg detection have been developed. Model prediction or satellite methods are often used to monitor large icebergs (width over 100 meters). All-weather remote sensing methods such as spaceborne polarimetric synthetic aperture radar (SAR) and airborne radar are mainly used to detect small and medium-sized icebergs (width over 10 meters) due to their resolution and revisit time. These iceberg targets are easier 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 during navigation. For these targets, the crew generally judges them based on the shipboard detection system combined with experience. The detection system includes ship radar, optical and infrared sensors. However, due to the extremely harsh sea conditions in the Arctic, the ship radar signal will be submerged by sea clutter on the radar image, making it difficult to distinguish it from clutter. Optical sensors are limited by fog and low light conditions, which will also greatly reduce the effectiveness of detection. Therefore, infrared iceberg detection is an important supplementary method for iceberg monitoring in Arctic waters under harsh marine conditions.
[0004] Infrared sensors can capture the contrast difference between icebergs and the surrounding seawater background, are highly immune to weather such as fog and haze, have the ability to work all day, and can work under poor lighting conditions. Infrared detection technology has been proven to be an effective tool for detecting icebergs. However, infrared iceberg target detection faces many challenges. First, as the detection distance increases, the imaging contrast between icebergs and seawater decreases, and low-contrast infrared iceberg data increases the difficulty of detection. In addition, the grayscale and local contrast of some strong waves and wakes are higher than the target, and they are easily misdetected as real targets. In addition, the shape of iceberg targets is varied and irregular, with little target pixel and texture information, and limited available features. At present, many traditional methods have been proposed for infrared small target detection technology, including filter-based methods, methods based on target and background feature calculation, and optimization methods through low-rank sparse matrix operations. However, these traditional methods often rely on manually extracted image features and pre-set hyperparameters, and they only have good performance in simple scenarios. Therefore, deep learning methods are gradually being 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 failure to detect targets. Therefore, a method of using small target semantic features combined with target edge information for target detection was proposed. This method makes up for the defect that target detection methods based on deep learning methods are prone to losing small target features. However, infrared iceberg images have complex scenes, low contrast, and severe sea clutter. Figure 1 As shown in the figure, the existing small target detection method cannot adapt to dim targets and strong sea clutter scenes, which will produce false alarms and misdetections, resulting in low accuracy of infrared iceberg target detection. Summary of the invention
[0005] The purpose of the present invention is to solve the problem of low accuracy of infrared iceberg target detection in existing target detection methods, and 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] Obtain 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;
[0008] The training set of the ADENet network is the original infrared iceberg image and the infrared iceberg image with the target position marked;
[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 obtain an infrared iceberg feature map set E={E1, E2, E3, E4, E5} consisting of five infrared iceberg image feature maps with different resolutions;
[0011] in, i∈{1,2,3,4,5},E i is the feature map of the ith infrared iceberg image, C i YesE i The number of channels, H i YesE i Height, W i YesE i Width;
[0012] The global contrast enhancement module uses the feature map E5 to extract the target global features and obtain the feature map and the guiding feature E guide ;
[0013] The gated semantic feature extraction module is used for target semantic features, using feature map E4 and feature E guide Extract feature map Using feature maps And feature map E3 to obtain feature map Using feature maps And feature map E2 to obtain feature map
[0014]
[0015] The detail feature extraction module uses the feature map E1, Extract shallow detail features of the target and obtain feature maps
[0016] The decoder is used to Decode and aggregate to obtain target recognition results in infrared iceberg images.
[0017] Furthermore, the feature map Obtained through:
[0018]
[0019] Among them, Conv() is a 3×3 convolution operation. is the intermediate feature map. Conv 1×1 () is a 1×1 convolution operation, Conv 3×3 () is a 3×3 convolution operation, Conv 5×5 () is a 5×5 convolution operation, Conv 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] Among them, SA() is the spatial attention operation.
[0023] Furthermore, the gated semantic feature extraction module is used to extract the target semantic feature using the feature map E4 and the feature E guide Extract feature map Using feature maps And feature map E3 to obtain feature map Using feature maps And feature map E2 to obtain feature map Specifically:
[0024]
[0025] Among them, up() is the upsampling operation, is the intermediate feature map, It is a point-by-point multiplication operation of the spatial dimension, and j can be 2, 3, or 4.
[0026] Further, Specifically:
[0027]
[0028]
[0029] Among them, Conv di1 It is a dilated convolution operation with Dilated Rate = 1, stack() is a feature map stacking operation, and mean() is a feature map average calculation operation. is the intermediate feature map, Conv di3 () is the dilated convolution operation with Dilated Rate = 3, Conv di5 () is the dilated convolution operation with Dilated Rate = 5, Conv di7 () is the dilated convolution operation with Dilated Rate = 7, x gatek is the feature of the output of the kth threshold mechanism operation, and k is 1, 3, or 5.
[0030] Furthermore, the characteristics of the output of the k-th threshold mechanism operation are specifically:
[0031]
[0032] Among them, gate() is the threshold mechanism operation, and Sigmoid() is the Sigmoid function.
[0033] Furthermore, the detail feature extraction module uses the feature map E1, Extract shallow detail features of the target and obtain feature maps Specifically:
[0034]
[0035] in, is the intermediate feature map, Aver() is the average pooling operation, Max() is the maximum pooling operation, CA() is the channel attention operation, is the feature map output by the gated semantic feature extraction module, and ⊙ is the point-by-point multiplication operation in the channel dimension.
[0036] Furthermore, the decoder is used to decode the feature map Decode and aggregate to obtain the target recognition results in the infrared iceberg image, specifically:
[0037] A1. Using feature graphs Get feature map D5:
[0038]
[0039] A2, initialize i=5;
[0040] A3. Using feature graphs and D i Get feature map D i-1 ;
[0041] A4. Determine whether 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 utilization characteristic diagram in A3 and D i Get feature map D i-1 , specifically:
[0043]
[0044] Among them, 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 loss 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 true value of the target in the infrared iceberg image.
[0048] The beneficial effects of the present invention are:
[0049] The present invention proposes an infrared iceberg target detection method (ADENet) based on an adaptive feature enhancement interactive network. The present invention uses a global contrast enhancement module to adaptively enhance low-contrast images with the target as the center, and restores the potential global features in the image. The feature guidance branch in the global contrast enhancement module provides guidance information for the extraction of target semantic features. The present invention adaptively enhances detail features such as target edges by designing a detail feature extraction module, and adaptively extracts target edge contours and enhances image detail features. The present invention avoids false alarms and misdetections caused by dim targets in infrared iceberg image fields and strong sea clutter scenes, and improves the accuracy of infrared iceberg target detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 It is a three-dimensional visual image of infrared iceberg targets and sea clutter;
[0051] Figure 2 Enhance the interactive network structure graph for adaptive features;
[0052] Figure 3 This is the structure diagram of the gated semantic feature extraction module;
[0053] Figure 4 This is the structure diagram of the global contrast enhancement module;
[0054] Figure 5 Calculate the schematic diagram based on the detail feature extraction module;
[0055] Figure 6(a) is a comparison of the detection effects of ADENet and other methods on the SIRST dataset;
[0056] Figure 6(b) is a comparison of the detection effects of ADENet and other methods on the SFIIT dataset. DETAILED DESCRIPTION
[0057] Specific implementation method 1: 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: obtaining the infrared iceberg image to be identified, inputting the infrared iceberg image to be identified into the trained ADENet network, and obtaining the target recognition result.
[0058] The training set of the ADENet network is the original infrared iceberg image and the infrared iceberg image with the target position marked;
[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, which is used to perform downsampling calculation on 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 five infrared iceberg image feature maps 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 is the feature map of the ith 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 YesE i The number of channels, H i YesE i Height, W i YesE i Width;
[0062] The global contrast enhancement module GCE uses the feature map E5 to extract the target global features and obtain the feature map and the guiding feature E guide ;
[0063] The gated semantic feature extraction module is used for target semantic features, using feature map E4 and feature E guide Extract feature map Using feature maps And feature map E3 to obtain feature map Using feature maps And feature map E2 to obtain feature map
[0064]
[0065] The detail feature extraction module DFE uses the feature map E1, Extract shallow detail features of the target and obtain feature maps
[0066] The decoder is used to Decode and aggregate to obtain target recognition results in infrared iceberg images.
[0067] Specific implementation method 2: 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 obtain the feature map The contrast enhancement branch includes in sequence: a first convolutional layer of 3×3, a second convolutional layer of 1×1, a third convolutional layer of 3×3, a fourth convolutional layer of 5×5, a fifth convolutional layer of 7×7, a first feature connection layer, a sixth convolutional layer of 3×3, a feature map Output layer;
[0069] The second convolution layer, the third convolution layer, the fourth convolution layer, and the fifth convolution layer are parallel convolution layers, and the input feature maps are all output feature maps of the first convolution layer;
[0070] The first 3×3 convolutional layer uses feature map E5 to obtain Specifically:
[0071]
[0072] The inputs of 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 obtained by connecting the feature map layer Specifically:
[0073]
[0074] Among them, Concate() is a feature connection operation;
[0075] The input of the sixth 3×3 convolutional layer is the feature map After the convolution operation, it is input into the feature map Output layer;
[0076] The feature map The output layer will be the convolution feature map and Sum the elements to get the feature map Specifically:
[0077]
[0078] Among them, Conv() is a 3×3 convolution operation;
[0079] The feature-guided branch uses the feature map and feature map Get the guiding feature E guide ;
[0080] The feature guidance branch includes: a first feature connection layer, a 3×3 first convolution layer, a spatial attention mechanism layer, and a 3×3 second convolution layer;
[0081] The feature guide branch, the specific processing process is as follows:
[0082]
[0083] Among them, E guide is the guiding feature and SA is the spatial attention operation.
[0084] In this implementation, E5 has the lowest resolution and does not have shallow detail features, but contains global information such as global contrast features and large-scale contours. 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, and information is obtained through parallel convolutions with convolution kernels of 1×1, 3×3, 5×5, and 7×7, and then the feature maps of the parallel convolutions are connected: Then, the output of the input and contrast enhancement branches is input into the feature guidance branch, and the image contrast features are emphasized through the spatial attention mechanism, and the guided feature E is output. guide , providing guidance information for the gated semantic feature extraction module. In this implementation, the first feature connection layer, the first convolution layer, and the second convolution layer of the macro are not the same convolution layer, but the first feature connection layer, the first convolution layer, and the second convolution layer of the corresponding branch.
[0085] Specific implementation method three: Figure 3 As shown, the gated semantic feature extraction module includes: feature map Get unit, feature map Get unit, feature map Get unit, feature map 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 Dilated Rate=1, and a first feature summation subunit;
[0087] The feature map E j Enter the first dilated convolution subunit with Dilated Rate = 1, and the E after dilated convolution j With feature E j The feature map is obtained by the first feature summation subunit Specifically:
[0088]
[0089] Among them, Convdi1 It is the first dilated convolution operation with Dilated Rate = 1;
[0090] The feature map The acquisition unit includes, in sequence: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolution layer with a Dilated Rate of 3, and a second feature summation subunit;
[0091] The threshold mechanism sub-unit includes: 3*3 convolution sub-layer, sigmoid function sub-layer, spatial dimension point-by-point multiplication operation sub-layer; the input of the 3*3 convolution sub-layer is the feature map output by the previous layer The output of the 3*3 convolution sublayer is the input of the sigmoid function sublayer. The output of the sigmoid function sublayer and the output of the previous feature summation layer are the input of the spatial dimension point-by-point multiplication operation sublayer. The output of the spatial dimension point-by-point multiplication operation sublayer is the output of the threshold mechanism subunit, which is:
[0092]
[0093] Among them, k is 1, 3, 5, Pick x gatek is the output feature of the threshold mechanism subunit, Sigmoid represents the Sigmoid function, It is a point-by-point multiplication operation in the spatial dimension, and gate() is a threshold mechanism operation;
[0094] Feature Map The first threshold mechanism subunit input in the acquisition unit is the feature map The output is x gate1 ; Output x of the first threshold mechanism subunit gate1 With the feature map E j The summation result is obtained through the first feature summation subunit, and then the summation result passes through the first dilated convolution layer with Dilated Rate = 3. The output of the first dilated convolution layer with Dilated Rate = 3 and the output of the second feature summation subunit are again passed through the feature summation subunit to obtain the feature map Specifically:
[0095]
[0096] Among them, Conv di3 () is the dilated convolution operation with Dilated Rate = 3;
[0097] Feature Map The acquisition unit includes: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolution layer with Dilated Rate = 5, and a second feature summation subunit in sequence; feature map Get layer processing flow and feature map The acquisition layer processing flow is similar, specifically:
[0098]
[0099] Among them, Conv di5 () is the dilated convolution operation with Dilated Rate = 5;
[0100] Feature Map The acquisition units include: a first threshold mechanism subunit, a first feature summation subunit, a first dilated convolution layer with Dilated Rate = 7, and a second feature summation subunit in sequence; feature map Get layer processing flow and feature map The acquisition layer processing flow is similar, specifically:
[0101]
[0102] Among them, Conv di7 () is the dilated convolution operation with Dilated Rate = 7;
[0103] The feature map stacking unit is used to stack the feature maps Perform a stacking operation to obtain a 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 a feature map Specifically:
[0105]
[0106] Among them, stack() is the feature map stacking operation, mean() is the operation of calculating the average value of the feature map;
[0107] Said The feature map output unit is used to Perform channel attention operations and spatial attention operations to obtain feature maps Then use the feature map Feature Map E j , feature map Get feature map Specifically:
[0108]
[0109]
[0110] Among them, up() is the upsampling operation.
[0111] In this implementation, we first obtain multi-scale information of the target at different dilation sizes through dilated convolution operations. We introduce a threshold mechanism to increase the information propagation channel between adjacent dilated convolution branches. The mechanism designs the feature map generated by the convolution layer and the Sigmoid function filter. Get unit, feature map Get unit, feature map Get unit, feature map The first dilated convolution subunit, the first feature summation subunit, the second feature summation subunit, and the first threshold mechanism subunit involved in the acquisition unit do not refer to the same subunit, but the subunits in their respective corresponding layers.
[0112] Specific implementation method four: Figure 5 As shown, the detail feature extraction module includes: an element separation unit, a feature map Get unit, feature map Get unit, feature map Get unit, feature map Get unit, feature map Get unit;
[0113] The element separation unit includes a first 3×3 convolutional layer;
[0114] Feature map E1 input input first 3 × 3 convolutional layer to obtain feature map Specifically:
[0115]
[0116] The feature map The acquisition unit includes: an upsampling layer, a point-by-point multiplication operation layer in spatial dimension, a spatial attention layer, and a feature map summation layer; the feature map Get the input of the unit as feature map and feature map First, after the upsampling layer, the resolution is the same as E1, and then after the upsampling and The input spatial dimension point-by-point multiplication operation layer is used as the input of the spatial attention layer. 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 Get subunits and feature maps Get subunits;
[0119] The feature map The acquisition sub-units include: channel attention layer, channel dimension point-by-point multiplication element layer, and feature map summation layer;
[0120] The input of the channel attention layer is It passes through the channel attention layer, the channel dimension point-by-point multiplication element layer, and the feature map summation layer in sequence. Get the feature map Specifically:
[0121]
[0122] Among them, ⊙ is the channel dimension point-wise multiplication operator, CA() is the channel attention operation;
[0123] The feature map The acquisition subunit includes: a spatial attention layer, a spatial dimension point-by-point multiplication operation layer, and a feature map summation layer; the feature map Get subunit input as feature map The feature map is obtained by sequentially passing through the spatial attention layer, the spatial dimension point-by-point multiplication operation layer, and the feature map summation layer. Specifically:
[0124]
[0125] The feature map The acquisition unit includes: feature map Get subunits and feature maps Get subunits;
[0126] The feature map The acquisition subunits include: channel attention layer, channel dimension point-by-point multiplication operation layer, feature map summation layer; feature map Get the input of the subunit as a feature map The feature map is obtained by sequentially passing through the channel attention layer, the channel dimension point-by-point multiplication operation layer, and the feature map summation layer. Specifically:
[0127]
[0128] The feature map The acquisition subunit includes: a spatial attention layer, a spatial dimension point-by-point multiplication operation layer, and a feature map summation layer; the feature map Get subunit input as feature map The feature map is obtained by sequentially passing through the spatial attention layer, the spatial dimension point-by-point multiplication operation layer, and the feature map summation layer. Specifically:
[0129]
[0130] The feature map The acquisition unit includes: average pooling layer, maximum pooling layer, and feature map subtraction operation layer;
[0131] The average pooling layer is used to Perform average pooling operation;
[0132] The maximum pooling layer is used to Perform average pooling operation;
[0133] The feature map subtraction operation layer performs a subtraction operation on the output of the average pooling layer and the output of the maximum pooling layer to obtain the feature map Specifically:
[0134]
[0135] Among them, Aver() is the average pooling operation, and Max() is the maximum pooling operation;
[0136] The feature map The acquisition unit includes: feature connection layer, Sigmoid activation function layer, and feature map summation layer;
[0137] The input of the feature map connection layer is the feature map After passing through the Sigmoid activation function layer and the feature map summation layer in sequence, the feature map is obtained Specifically:
[0138]
[0139] Among them, Concate() is the feature map connection operation.
[0140] In this step, average pooling and maximum pooling are applied respectively, and then they are subtracted to adaptively perceive shallow detail information such as edges, without the need for separate supervision of edge GT. Then, the feature map after subtraction and the two feature maps before subtraction are connected using skip connections to form the output of the detail feature extraction module. Finally, we use the difference feature map And the feature maps of the two branches Connect them, use 1×1 convolution to adaptively aggregate the features, and finally add them to the input shallow feature map E1 to obtain the output of the detail feature extraction module DFE.
[0141] The characteristic map in this embodiment Get subunits and feature maps Get subunits and feature maps Get subunits and feature maps The channel attention layer, channel dimension point-by-point multiplication element layer, feature map summation layer, spatial attention layer, and spatial dimension point-by-point multiplication operation layer mentioned in the acquisition sub-unit are not the same layer, but the channel attention layer, channel dimension point-by-point multiplication element layer, feature map summation layer, spatial attention layer, and spatial dimension point-by-point multiplication operation layer in their respective corresponding sub-units.
[0142] Specific implementation method 5: The decoder is used to Decode and aggregate to obtain the target recognition results in the infrared iceberg image, specifically:
[0143] A1. Using feature graphs Get feature map D5:
[0144]
[0145] A2, initialize i=5;
[0146] A3. Using feature graphs and D i Get feature map D i-1 :
[0147]
[0148] Among them, BN() is the batch normalization operation, and ReLU() is the ReLU activation function;
[0149] A4. Determine whether 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 6: The ADENet network training adopts the following mixed loss function loss 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 true value of the target in the infrared iceberg image.
[0153] Example: In order to verify the beneficial effects of the present invention, the ADENet network proposed in the present invention is compared with the existing network. The results are shown in Table 1 and the effect comparison diagram is shown in Figure 6(a)-Figure 6(b) As shown;
[0154] Table 1
[0155]
[0156] According to the above table, as shown in Figure 6(a), on the public infrared small target dataset SIRST, ADENet has the best evaluation indicators in nIoU, Fa, and Prec, and has better small target segmentation capabilities than other methods. Due to the design of the contrast enhancement module in the ADENet network, the ADENet method reduces the false alarm of target detection and obtains a higher F1 score. As shown in Figure 6(b), on the infrared iceberg target dataset SFIIT, the ADENet method verifies the superiority of the network in iceberg target detection in all aspects of data. The IoU and nIoU indicators are much higher than other methods, and a balance is achieved in the Prec and Rec indicators. It can detect iceberg targets under the condition of better control of false alarms and obtain the best F1 score.
Claims
1. Infrared iceberg target detection method based on adaptive feature enhancement interactive network, Features The specific process of the method is: obtaining an infrared iceberg image to be identified, inputting the infrared iceberg image to be identified into a trained ADENet network, and obtaining a target recognition result; The training set of the ADENet network is the original infrared iceberg image and the infrared iceberg image with the target position marked; 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 obtain an infrared iceberg feature map set E={E 1 ,E 2 ,E 3 ,E 4 ,E 5 }; in, i∈{1,2,3,4,5},E i is the feature map of the ith infrared iceberg image, C i YesE i The number of channels, H i YesE i Height, W i YesE i Width; The global contrast enhancement module uses the feature map E 5 Extract the global features of the target and obtain the feature map and the guiding feature E guide ; The gated semantic feature extraction module is used for target semantic features, using the feature map E 4 and feature E guide Extract feature map Using feature maps And feature map E 3 Get feature map Using feature maps And feature map E 2 Get feature map The detail feature extraction module uses the feature map E 1 , Extract shallow detail features of the target and obtain feature maps The decoder is used to Decode and aggregate to obtain target recognition results in infrared iceberg images.
2. According to claim 1, the infrared iceberg target detection method based on adaptive feature enhancement interactive network, Features: The feature map Obtained through: Among them, Conv() is a 3×3 convolution operation. is the intermediate feature map. Conv 1×1 () is a 1×1 convolution operation, Conv 3×3 () is a 3×3 convolution operation, Conv 5×5 () is a 5×5 convolution operation, Conv 7×7 () is a 7×7 convolution operation, and Concate() is a feature connection operation.
3. The infrared iceberg target detection method based on the adaptive feature enhancement interactive network according to claim 2, Features: The guiding feature E guide , specifically: Among them, SA() is the spatial attention operation.
4. The infrared iceberg target detection method based on the adaptive feature enhancement interactive network according to claim 3, Features: The gated semantic feature extraction module is used to extract the target semantic features using the feature map E 4 and feature E guide Extract feature map Using feature maps And feature map E 3 Get feature map Using feature maps And feature map E 2 Get feature map Specifically: Among them, up() is the upsampling operation, is the intermediate feature map, It is a point-by-point multiplication operation of the spatial dimension, and j can be 2, 3, or 4.
5. The infrared iceberg target detection method based on the adaptive feature enhancement interactive network according to claim 4, Features: Specifically: Among them, Conv di1 It is a dilated convolution operation with Dilated Rate = 1, stack() is a feature map stacking operation, and mean() is a feature map average calculation operation. is the intermediate feature map, Conv di3 () is the dilated convolution operation with DilatedRate=3, Conv di5 () is the dilated convolution operation with Dilated Rate = 5, Conv di7 () is the dilated convolution operation with Dilated Rate = 7, x gatek is the feature of the output of the kth threshold mechanism operation, and k is 1, 3, or 5.
6. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 5, Features: The characteristics of the k-th threshold mechanism operation output are specifically: Among them, gate() is the threshold mechanism operation, and Sigmoid() is the Sigmoid function.
7. The infrared iceberg target detection method based on the adaptive feature enhancement interactive network according to claim 6, Features: The detail feature extraction module uses the feature map E 1 , Extract shallow detail features of the target and obtain feature maps Specifically: in, is the intermediate feature map, Aver() is the average pooling operation, Max() is the maximum pooling operation, CA() is the channel attention operation, is the feature map output by the gated semantic feature extraction module, and ⊙ is the point-by-point multiplication operation in the channel dimension.
8. The infrared iceberg target detection method based on the adaptive feature enhancement interactive network according to claim 7, Features: The decoder is used to Decode and aggregate to obtain the target recognition results in the infrared iceberg image, specifically: A1. Using feature graphs Get feature map D 5 : Among them, BN() is the batch normalization operation, and ReLU() is the ReLU activation function; A2, initialize the characteristic map number i = 5; A3. Using feature graphs and D i Get feature map D i-1 ; A4. Determine whether i is 2. If i=2, output the target recognition result D in the infrared iceberg image. 1 , 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, Features: The utilization characteristic diagram of A3 and D i Get feature map D i-1 , specifically: Among them, ReLU() is the ReLU activation function.
10. The infrared iceberg target detection method based on adaptive feature enhancement interactive network according to claim 9, Features: The loss function of the ADENet network is a hybrid loss function loss 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 true value of the target in the infrared iceberg image.
Citation Information
Patent Citations
Night airport net airspace infrared small target detection method, system and application
CN115410012A
Infrared small target detection method based on attention-oriented pyramid fusion
CN116071676A
Infrared weak and small target detection method based on global information target enhancement
CN116883912A
Method for training shallow convolutional neural networks for infrared target detection using a two-phase learning strategy
US20230237788A1
Cited By
Infrared iceberg target detection method based on global feature and multi-resolution semantic feature interaction network
CN119762760A
