A synthetic aperture radar image target detection method, device and medium

Abstract

The application discloses a synthetic aperture radar image target detection method, equipment and medium, and belongs to the technical field of SAR image target detection. The synthetic aperture radar image target detection method comprises the following steps: S1: inputting a SAR image; S2: constructing a feature separation module; S3: performing feature extraction; S4: constructing an adaptive feature aggregation network; S5: obtaining a target detection result; and S6: optimizing model parameters. The application solves the problem that the prior art is prone to causing high false detection rate and missed detection rate of multi-scale target detection. The application separates noise features and adaptively aggregates multi-scale features, reduces the interference of a complex background, improves the influence of a noise environment and multi-scale targets on detection, directly restricts the position and aspect ratio of a prediction box by using a regression loss, optimizes the multi-scale target detection effect of a model, reduces the false detection rate and missed detection rate of SAR image target detection, and improves the robustness of the model.

Description

Technical Field

[0001] This invention relates to the field of SAR image target detection technology, specifically to a synthetic aperture radar image target detection method, device, and medium. Background Technology

[0002] Synthetic Aperture Radar (SAR) is an active imaging observation system that images objects by actively emitting microwaves and receiving reflected signals. SAR is capable of operating in all weather conditions and at all times, and is widely used in military, agriculture, forestry, geology, and marine environmental fields. Target detection and recognition is a crucial problem in SAR image processing, possessing significant theoretical research value and practical application value.

[0003] Currently, traditional and deep learning methods for target detection in SAR images achieve high accuracy and low false negative rates for most SAR image scenes. However, due to the unique noise interference, complex backgrounds, and multi-scale targets inherent in SAR images, the false positive and false negative rates for multi-scale target detection are prone to be high, thus failing to meet current requirements. To address this, we propose a synthetic aperture radar (SAR) image target detection method, device, and medium. Summary of the Invention

[0004] The purpose of this invention is to provide a synthetic aperture radar (SAR) image target detection method, device, and medium. By separating noise features and adaptively aggregating multi-scale features, the invention improves the impact of noise environment and multi-scale targets on detection, reduces the false detection rate and false negative rate of SAR image target detection, improves the robustness of the model, and solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a synthetic aperture radar image target detection method, comprising the following steps:

[0006] S1: Input SAR image;

[0007] S2: Construct a feature separation module to process the input SAR image, suppress noise interference, and enhance image feature representation;

[0008] S3: Select a multi-scale feature extraction network and input the image processed by the feature separation module into the multi-scale feature extraction network for feature extraction.

[0009] S4: Construct an adaptive feature aggregation network. Input the feature map obtained from the feature extraction network into the adaptive feature aggregation network. Generate an enhanced semantic feature map by fusing it with a special dual-branch structure and dynamic weights.

[0010] S5: The feature map output by the adaptive feature aggregation network will be fed into... A classification detector composed of convolutional layers obtains the target detection results;

[0011] S6: The regression loss function is used, combined with the classification loss function and the confidence loss function, and the model parameters are optimized through the backpropagation algorithm.

[0012] Preferably, the calculation method of the feature separation module is as follows:

[0013]

[0014] in, , represents the input image, Indicates to The feature map obtained after channel dimension expansion Indicates to The feature map obtained after channel weighting This represents the output of the feature separation module. This indicates a morphological etching operation. express convolution, Indicates global average pooling. Indicates a fully connected layer. This indicates multiplication by channel dimension.

[0015]

[0016] in, This represents the activation function of the module. This represents the value of a specific channel and spatial location within the input features. The output value corresponding to this position after function mapping.

[0017] Preferably, the calculation method of the multi-scale feature extraction network is as follows:

[0018]

[0019]

[0020] in, The output of the feature extraction network represents the first... Layer feature map, Indicates the module's input. and These represent the convolution kernel as follows: Convolution and convolution, Represents a non-linear activation function. This represents the batch normalization function;

[0021]

[0022] in, This indicates max pooling with a kernel size of 3. This indicates that the inputs will be concatenated along the channel dimension.

[0023] Preferably, the adaptive feature aggregation network is calculated as follows:

[0024]

[0025] in, , and These represent the outputs of aggregation modules at different output resolutions in the adaptive feature aggregation network, namely the semantic feature maps enhanced at small, medium, and large resolutions, respectively. arrive , arrive and arrive These represent the weighting coefficients in different aggregation modules. This represents a special type of two-branch structure. This means upsampling the feature map to the same size as the output feature map.

[0026] Preferably, the specific components of the special double-branch structure include:

[0027]

[0028] in, Indicates the location branch, Indicates the downsampling branch. Represents the input feature map, This indicates taking the average. express Activation function Right now function, This represents a two-branch structure, and uses... Second-rate Convolutional downsampling and Correspondence indicates the use of Convolution Sub-sampling; This means that the feature map is downsampled to the same size as the output feature map through max pooling.

[0029] Preferably, the formula for calculating the regression loss is as follows:

[0030]

[0031] = +

[0032]

[0033] in, This measure represents the comprehensive geometric difference between the ground truth bounding box and the predicted bounding box, reflecting the degree of deviation of the predicted bounding box in both location coordinates and size scale. This represents the square of the distance between the center points of the ground truth bounding box and the center point of the predicted bounding box. and These represent the width and height of the actual bounding box, respectively. and Indicates the width and height of the prediction box; This represents the trade-off factor for balancing different aspect ratios. This represents the square of the diagonal length of the actual bounding box. This represents the square of the distance between the top-left corners of the ground truth bounding box and the predicted bounding box. This represents the square of the distance between the bottom right corners of the ground truth bounding box and the predicted bounding box. This represents the square of the diagonal length of the rectangles containing the ground truth bounding box and the predicted bounding box.

[0034] Preferably, the formula for calculating the classification loss is as follows:

[0035]

[0036] in, Represents classification loss, Indicates the total number of categories. express Activation function This represents the true value of the current category. This represents the predicted value for the current category.

[0037] Preferably, the formula for calculating the confidence loss is as follows:

[0038]

[0039] in, This represents the confidence loss, where This indicates the total number of prediction boxes. Indicates the first One prediction box, Indicates the first The true label of each predicted bounding box Indicates the first The predicted value of each prediction box.

[0040] An electronic device includes a processor and a memory; the memory is used to store a program; the processor executes the program to implement a synthetic aperture radar image target detection method.

[0041] A computer-readable storage medium includes computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform a synthetic aperture radar image target detection method.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention utilizes the noise decoupling capability of the feature separation module, the multi-level feature integration capability of the adaptive feature aggregation network, and the prediction box optimization capability of the regression loss to improve the model's ability to suppress noise and accurately locate targets. This optimizes the detection accuracy of SAR image target detection in complex scenes, ensuring robustness and accuracy in the face of complex scenes and multi-scale targets. By separating noise features and adaptively aggregating multi-scale features, the impact of noisy environment and multi-scale targets on detection is reduced, interference from complex backgrounds is decreased, the false detection rate and false negative rate of SAR image target detection are reduced, and the robustness of the model is improved. Attached Figure Description

[0044] Figure 1 This is an image from the SAR image target detection dataset;

[0045] Figure 2 This is a network structure diagram of the feature separation module;

[0046] Figure 3 Diagram of the multi-scale feature extraction network structure;

[0047] Figure 4 Diagram of the adaptive feature aggregation network structure;

[0048] Figure 5 This is a diagram showing the target prediction results for a SAR image.

[0049] Figure 6 This is a flowchart of a synthetic aperture radar image target detection method according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To address the issue of high false positive and false negative rates in multi-scale target detection caused by noise interference, complex backgrounds, and the presence of multi-scale targets inherent in SAR images in existing technologies, please refer to [link to relevant documentation]. Figures 1-6 This embodiment provides the following technical solution:

[0052] A method for target detection in synthetic aperture radar images includes the following steps:

[0053] S1: Select an image from the SAR image target detection dataset and input the SAR image;

[0054] S2: Construct a feature separation module, which includes morphological erosion operation and channel attention mechanism. The input SAR image is processed by the feature separation module. The feature separation module dynamically generates multi-channel feature maps, suppresses noise interference and enhances image feature expression, so that the network can separate noise and target features.

[0055] S3: Select a multi-scale feature extraction network. Input the image processed by the feature separation module into the multi-scale feature extraction network for feature extraction. This network consists of a series of convolutional layers and can extract feature maps at different levels.

[0056] S4: Construct an adaptive feature aggregation network. Input the feature map obtained from the feature extraction network into the adaptive feature aggregation network, and generate an enhanced semantic feature map through a special dual-branch structure and dynamic weight fusion. This dual-branch structure and dynamic weight adjustment mechanism make the network focus more on the target region;

[0057] S5: The feature map output by the adaptive feature aggregation network will be fed into... The classification detector, composed of convolutional layers, obtains target detection results, which are then outlined in the SAR image. Different colored target boxes represent different types of predicted targets.

[0058] S6: A regression loss function is adopted, combined with a classification loss function and a confidence loss function. The model parameters are optimized through the backpropagation algorithm. In order to further optimize the model parameters, a regression loss function is introduced to optimize the position and aspect ratio of the predicted box, thereby further enhancing the accurate positioning capability. The model is also optimized through the backpropagation algorithm in combination with the classification loss function.

[0059] The calculation method for the feature separation module is as follows:

[0060]

[0061] in, , represents the input image, Indicates to The feature map obtained after channel dimension expansion Indicates to The feature map obtained after channel weighting This represents the output of the feature separation module. This indicates a morphological etching operation. express Convolution, the first convolution expands the number of channels to 64, that is... The second convolution adjusts the number of channels to 3, that is... , Indicates global average pooling. Indicates a fully connected layer. This indicates multiplication by channel dimension.

[0062]

[0063] in, This represents the activation function of the module. This represents the value of a specific channel and spatial location within the input features. The output value corresponding to this position after function mapping.

[0064] The computational method for multi-scale feature extraction networks is as follows:

[0065]

[0066]

[0067] in, The output of the feature extraction network represents the first... Layer feature map, Indicates the module's input. and These represent the convolution kernel as follows: Convolution and convolution, Represents a non-linear activation function. This represents the batch normalization function;

[0068]

[0069] in, This indicates max pooling with a kernel size of 3. This indicates that the inputs will be concatenated along the channel dimension.

[0070] The process of adaptive feature aggregation network, and the calculation method are as follows:

[0071]

[0072] in, , and These represent the outputs of aggregation modules at different output resolutions in the adaptive feature aggregation network, namely the semantic feature maps enhanced at small, medium, and large resolutions, respectively. arrive , arrive and arrive These represent the weighting coefficients in different aggregation modules. This represents a special type of two-branch structure. This means upsampling the feature map to the same size as the output feature map.

[0073] The specific components of the special double-branch structure include:

[0074]

[0075] in, Indicates the location branch, Indicates the downsampling branch. Represents the input feature map, This indicates taking the average. express Activation function Right now function, This represents a two-branch structure, and uses... Second-rate Convolutional downsampling and Correspondence indicates the use of Convolution Sub-sampling; This means that the feature map is downsampled to the same size as the output feature map through max pooling.

[0076] The formula for calculating regression loss is as follows:

[0077]

[0078] = +

[0079]

[0080] in, This measure represents the comprehensive geometric difference between the ground truth bounding box and the predicted bounding box, reflecting the degree of deviation of the predicted bounding box in both location coordinates and size scale. This represents the square of the distance between the center points of the ground truth bounding box and the center point of the predicted bounding box. and These represent the width and height of the actual bounding box, respectively. and Indicates the width and height of the prediction box; This represents the trade-off factor for balancing different aspect ratios. This represents the square of the diagonal length of the actual bounding box. This represents the square of the distance between the top-left corners of the ground truth bounding box and the predicted bounding box. This represents the square of the distance between the bottom right corners of the ground truth bounding box and the predicted bounding box. This represents the square of the diagonal length of the rectangles containing the ground truth bounding box and the predicted bounding box.

[0081] The formula for calculating classification loss is as follows:

[0082]

[0083] in, Represents classification loss, Indicates the total number of categories. express Activation function This represents the true value of the current category. This represents the predicted value for the current category.

[0084] The formula for calculating confidence loss is as follows:

[0085]

[0086] in, This represents the confidence loss, where This indicates the total number of prediction boxes. Indicates the first One prediction box, Indicates the first The true label of each predicted bounding box Indicates the first The predicted value of each prediction box.

[0087] An electronic device includes a processor and a memory; the memory is used to store a program; the processor executes the program to implement a synthetic aperture radar image target detection method.

[0088] A computer-readable storage medium includes computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform a synthetic aperture radar image target detection method.

[0089] In summary, the synthetic aperture radar (SAR) image target detection method, device, and medium of this invention first process the input SAR image through a feature separation module. This module includes morphological erosion operations and a channel attention mechanism to dynamically generate multi-channel feature maps, suppressing noise interference and enhancing image feature representation. Next, the processed image is fed into a multi-scale feature extraction network, which consists of a series of convolutional layers capable of extracting feature maps at different levels. Then, an adaptive feature aggregation network is used to dynamically fuse these multi-scale feature maps across layers, generating three enhanced semantic feature maps with different resolutions. This dynamic weight adjustment mechanism focuses more on the target region. Finally, to further optimize the model parameters… A regression loss function is introduced to optimize the position and aspect ratio of the predicted bounding box, further enhancing the accuracy of localization. The model is then optimized using a backpropagation algorithm in conjunction with a classification loss function. The noise decoupling capability of the feature separation module, the multi-level feature integration capability of the adaptive feature aggregation network, and the predicted bounding box optimization capability of the regression loss are utilized to improve the model's ability to suppress noise and accurately locate targets. This optimizes the detection accuracy of SAR image targets in complex scenes, ensuring robustness and accuracy in complex scenarios and with multi-scale targets. By separating noise features and adaptively aggregating multi-scale features, the impact of noisy environments and multi-scale targets on detection is mitigated, reducing the false positive and false negative rates of SAR image target detection and improving the model's robustness.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for target detection in synthetic aperture radar images, characterized in that, Includes the following steps: S1: Input SAR image; S2: Construct a feature separation module to process the input SAR image, suppress noise interference, and enhance image feature representation; The calculation method of the feature separation module is as follows: in, , represents the input image, Indicates to The feature map obtained after channel dimension expansion Indicates to The feature map obtained after channel weighting This represents the output of the feature separation module. This indicates a morphological etching operation. express convolution, Indicates global average pooling. Indicates a fully connected layer. This indicates multiplication by channel dimension; in, This represents the activation function of the module. This represents the value of a specific channel and spatial location within the input features. The output value corresponding to this position after function mapping; S3: Select a multi-scale feature extraction network and input the image processed by the feature separation module into the multi-scale feature extraction network for feature extraction. S4: Construct an adaptive feature aggregation network. Input the feature map obtained from the feature extraction network into the adaptive feature aggregation network. Generate an enhanced semantic feature map by fusing it with a special dual-branch structure and dynamic weights. The process of the adaptive feature aggregation network is calculated as follows: in, , and These represent the outputs of aggregation modules at different resolutions in the adaptive feature aggregation network, namely, the semantic feature maps enhanced at small, medium, and large resolutions, respectively. arrive , arrive and arrive These represent the weighting coefficients in different aggregation modules. This represents a special type of two-branch structure. This means upsampling the feature map to the same size as the output feature map; The specific components of the special double-branch structure include: in, Indicates the location branch, Indicates the downsampling branch. Represents the input feature map, This indicates taking the average. express Activation function Right now function, This represents a two-branch structure, and uses... Second-rate Convolutional downsampling and Correspondence indicates the use of Convolution Sub-sampling; This means downsampling the feature map to the same size as the output feature map using max pooling; S5: The feature map output by the adaptive feature aggregation network will be fed into... A classification detector composed of convolutional layers obtains the target detection results; S6: The regression loss function is used, combined with the classification loss function and the confidence loss function, and the model parameters are optimized through the backpropagation algorithm.

2. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The calculation method of the multi-scale feature extraction network is as follows: in, The output of the feature extraction network represents the first... Layer feature map, Indicates the module's input. and These represent the convolution kernel as follows: Convolution and convolution, Represents a non-linear activation function. This represents the batch normalization function; in, This indicates max pooling with a kernel size of 3. This indicates that the inputs will be concatenated along the channel dimension.

3. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The formula for calculating the regression loss is as follows: = + in, This measure represents the comprehensive geometric difference between the ground truth bounding box and the predicted bounding box, reflecting the degree of deviation of the predicted bounding box in both location coordinates and size scale. This represents the square of the distance between the center points of the ground truth bounding box and the center point of the predicted bounding box. and These represent the width and height of the actual bounding box, respectively. and This indicates the width and height of the prediction box. This represents the trade-off factor for balancing different aspect ratios. This represents the square of the diagonal length of the actual bounding box. This represents the square of the distance between the top-left corners of the ground truth bounding box and the predicted bounding box. This represents the square of the distance between the bottom right corners of the ground truth bounding box and the predicted bounding box. This represents the square of the diagonal length of the rectangles containing the ground truth bounding box and the predicted bounding box.

4. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The formula for calculating the classification loss is as follows: in, Represents classification loss, Indicates the total number of categories. express Activation function This represents the true value of the current category. This represents the predicted value for the current category.

5. The synthetic aperture radar image target detection method according to claim 1, characterized in that, The formula for calculating the confidence loss is as follows: in, This represents the confidence loss, where This indicates the total number of prediction boxes. Indicates the first One prediction box, Indicates the first The true label of each predicted bounding box Indicates the first The predicted value of each prediction box.

6. An electronic device, characterized in that, It includes a processor and a memory; the memory is used to store a program; the processor executes the program to implement a synthetic aperture radar image target detection method as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, Includes computer-executable instructions, which, when executed by one or more processors, cause the one or more processors to perform the synthetic aperture radar image target detection method according to any one of claims 1-5.

Images