Space small target detection method fusing microwave infrared information

Abstract

The invention discloses a space small target detection method fusing microwave infrared information, and belongs to the technical field of image processing. The invention discloses a space small target detection method fusing microwave infrared information, and the method comprises the steps: carrying out the radar data preprocessing, carrying out the matched filtering and Doppler processing of a radar target echo, generating a distance-Doppler graph, processing the distance-Doppler graph, and forming an RDPA graph; performing target shape feature extraction and analysis on the infrared image based on an image segmentation algorithm of threshold segmentation; and designing a space small target detection network fusing RDPA map and infrared map information. According to the space small target detection method fusing the microwave and infrared information, the detection information increment between different sensors is further mined by designing the space small target detection network, deeply fusing the detection information of the microwave sensor and the detection information of the infrared sensor, and expanding the target detection dimension; and the fine detection of the small space target is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for detecting small space targets by fusing microwave and infrared information. Background Art

[0002] At present, the detection of small targets in space is mainly based on infrared information or radar information. The infrared image-based target detection technology is mainly divided into two types: traditional target detection method and deep learning-based target detection method.

[0003] Traditional infrared small target detection algorithms are mainly divided into two different categories according to the different orders of detecting and tracking the target. They are the algorithms that detect and then track the target in a single frame image and the algorithms that track and then detect the target in a sequence of images. Target detection methods based on deep learning include methods based on target detection network models, generative adversarial networks, and semantic segmentation networks. However, current methods generally have limitations such as insufficient robustness to noise, too much prior information required, and being greatly affected by low signal-to-noise ratio and lack of shape and texture features.

[0004] Small target detection methods based on radar information mainly focus on traditional hardware upgrades and transmission waveform changes. However, upgrading hardware equipment usually costs a lot of money, and the change of signal waveform cannot fundamentally solve the problem of low signal clutter noise of echo signals. Summary of the invention

[0005] The purpose of the present invention is to provide a method for detecting small space targets by integrating microwave infrared information. By designing a small space target detection network that integrates RDPA image and infrared image information, the microwave sensor detection information and the infrared sensor detection information are deeply integrated, the target detection dimension is expanded, and the detection information increment between different sensors is further mined to achieve refined detection of small space targets.

[0006] The deep fusion of microwave sensor detection information and infrared sensor detection information expands the target detection dimension, further explores the detection information increment between different sensors, and realizes the refined detection of small targets in space.

[0007] To achieve the above object, the present invention provides a method for detecting small space targets by integrating microwave infrared information, comprising the following steps: S1, radar data preprocessing, radar target echo is processed by matched filtering and Doppler processing to generate a range-Doppler (RD) map, and the range-Doppler map is processed to form an RDPA map, which is a range-Doppler peak area map; S2, image segmentation algorithm based on threshold segmentation to extract and analyze target shape features of infrared images; S3. Design a spatial small target detection network that integrates RDPA image and infrared image information.

[0008] Preferably, in S1, the distance-Doppler map formula is: ; in represents convolution, represents a known transmitted signal, is the time reversal of the transmitted signal, represents the filter designed to enhance the signal-to-noise ratio of the target signal; K is the normalization factor, ensuring that the signal power remains within the appropriate range. represents the output of the matched filter, Indicates the received echo signal. represents the first Pulses, t Indicates time, is the fast Fourier transform of the slow time dimension, represents the result of coherent accumulation, represents the Doppler frequency, represents the target speed, Indicates wavelength.

[0009] Preferably, in S1, the range-Doppler image is processed as follows: in the range-Doppler image, the highest point of the peak is taken as the center point, and a square area around the center point is intercepted according to a side length d to form an RDPA image.

[0010] Preferably, in S2, the image segmentation algorithm specifically includes the following steps: S21, input image Convert to grayscale ; S22, determine the image extraction area threshold and segmentation threshold ; S23, grayscale image Convert to binary image ; ;

[0011] S24, remove binary images Less than the extraction area threshold The connected domain of S25, mark greater than the segmentation threshold The connected domain of S26, return greater than the segmentation threshold The starting position and row / column width of the connected component.

[0012] Preferably, in S2, when extracting shape features, invariant moments, length and width in the shape features are selected as metrics, and the Hu invariant moment method is used to calculate 7 Hu invariant moments of the segmented area to construct a shape feature vector.

[0013] Preferably, in S3, the spatial small target detection network includes an RDPA graph convolution module for extracting spectral features in the RDPA graph, an infrared convolution module for extracting background features in the infrared graph, and a fusion module for fusing the spectral features with the background features, and the fusion module sends the fused joint features to the fully connected layer for classification; the RDPA graph is used as the input of the RDPA graph convolution module, and the infrared image shape feature vector is used as the input of the infrared convolution module.

[0014] Preferably, in S3, the RDPA graph convolution module is provided with an attention module for weighting spectral features and a fusion module for fusing spectral features at different levels.

[0015] Preferably, in S3, the infrared convolution module is provided with an attention module for weighting background features and a fusion module for fusing background features at different levels.

[0016] Preferably, the formula of the attention module is: ; in, represents the input features, Represents the output attention mechanism characteristics; Represents the convolution layer. The number of channels and dimensions of the feature remain unchanged after passing through this convolution layer. Responsible for generating attention weights at each feature point and comparing them with the input feature Multiply and exert attention impact.

[0017] Preferably, the formula of the fusion module is: ; in, and represents the input of the fusion module, is the output of the fusion module, Describes the connection of the input feature map in the channel dimension, represents the batch normalization operation, Represents the activation function.

[0018] The method for detecting small space targets by integrating microwave infrared information described in the present invention designs a small space target detection network integrating RDPA image and infrared image information, deeply integrates the detection information of microwave sensors and the detection information of infrared sensors, expands the target detection dimension, further mines the detection information increments between different sensors, and realizes the refined detection of small space targets.

[0019] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a method for detecting small space targets by integrating microwave infrared information according to the present invention; Figure 2 It is the network architecture of the space small target detection of the present invention; Figure 3 This is a diagram showing the effect of detecting small space targets according to the present invention. DETAILED DESCRIPTION

[0021] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] like Figure 1 A method for detecting small space targets by integrating microwave infrared information comprises the following steps: S1. Radar data preprocessing.

[0023] The radar target echo is processed by matched filtering and Doppler processing to generate a range-Doppler (RD) map. The range-Doppler map is processed to form an RDPA map.

[0024] The range-Doppler map formula is: ; in represents convolution, Indicates the transmission signal. is the time reversal of the transmitted signal, represents the designed filter, K is the normalization factor, represents the output of the matched filter, Indicates the received echo signal. represents the first Pulses, t Indicates time, is the fast Fourier transform of the slow time dimension, represents the result of coherent accumulation, represents the Doppler frequency, represents the target speed, Indicates wavelength.

[0025] The external features of the target are different, and the amplitude of the protrusion and the energy distribution around it are different in the RD map. However, the background clutter information occupies a large area in the RD map and blocks the useful information. In order to improve the information density of the training sample, the highest point of a peak is taken as the center point, and the fixed side length is used to calculate the information density of the training sample. d Cut off the square area around the center point to form the RDPA graph. d The size of is obtained from data statistics. Different types of targets produce different scales of convex peaks. Based on this, the radar target detection problem is simplified to the RD peak area (RDPA) map classification problem.

[0026] S2. Image segmentation algorithm based on threshold segmentation is used to extract and analyze target shape features of infrared images.

[0027] The image segmentation algorithm specifically includes the following steps: S21, input image Convert to grayscale ; S22, determine the image extraction area threshold and segmentation threshold ; S23, grayscale image Convert to binary image ; ;

[0028] S24, remove binary images Less than the extraction area threshold Thr1 The connected domain of S25, mark greater than the segmentation threshold The connected domain of S26, return greater than the segmentation threshold The starting position and row / column width of the connected component.

[0029] When extracting regional shape features after image segmentation, the invariant moments, length and width in the shape features are selected as metrics, and the Hu invariant moment method is used to calculate the 7 Hu invariant moments of the segmented region to construct the shape feature vector.

[0030] S3. Design a spatial small target detection network that integrates RDPA image and infrared image information.

[0031] like Figure 2As shown in the figure. The spatial small target detection network includes an RDPA graph convolution module for extracting spectral features from the RDPA graph, an infrared convolution module for extracting background features from the infrared graph, and a fusion module for fusing spectral features with background features. The fusion module sends the fused joint features to the fully connected layer for classification. The RDPA graph is used as the input of the RDPA graph convolution module, and the infrared image shape feature vector is used as the input of the infrared convolution module.

[0032] The network uses CNN and residual structure, and consists of two parallel feature extraction branches. One is responsible for extracting the spectral features of the peak area in the RD map. The other is used to extract the background features contained in the infrared image. The purpose of the convolution layer is to extract features from the connected multi-channel feature maps and perform channel dimension reduction by controlling the number of convolution kernels. The attention module will act on these features to obtain features weighted by the attention mechanism, and fuse them with the backbone network through the fusion module. The attention mechanism can help the model ignore irrelevant information, thereby improving detection performance.

[0033] The formula of the attention module is: ; in, represents the input features, Represents the output attention mechanism characteristics; Represents the convolution layer. The number of channels and dimensions of the feature remain unchanged after passing through this convolution layer. Responsible for generating attention weights at each feature point and comparing them with the input feature Multiply and exert attention impact.

[0034] The features are input into the convolution and residual structures to further extract deep features. The appropriate residual structure integrates the shallow feature information into the deep network structure, so that some key features of the input data will not be gradually lost as the number of network layers increases. At the same time, the residual structure can also alleviate the gradient disappearance and explosion that are easily caused by the deepening of the network. Different from the ordinary residual structure, when the residual branch features are integrated into the backbone network, the fusion module is used to avoid the destruction of feature integrity caused by direct addition.

[0035] The formula of the fusion module is as follows, which mainly includes a connection operation and a convolution block.

[0036] ; in, and denotes the input of the fusion module, which have the same size but different contents. is the output of the fusion module, Describes the connection of the input feature map in the channel dimension, Define the convolution layer, the number of convolution kernels is equal to the input features or The number of channels and the size of the convolution kernel are 3×3. represents the batch normalization operation, Represents the activation function.

[0037] Finally, the key features obtained from the RD image and infrared image branches are fused through the fusion module, and these key features are flattened into a one-dimensional vector and sent to the fully connected layer for classification. Figure 3 shown.

[0038] Therefore, the method for detecting small space targets by integrating microwave infrared information described in the present invention is adopted. By designing a small space target detection network that integrates RDPA image and infrared image information, the microwave sensor detection information and the infrared sensor detection information are deeply integrated, the target detection dimension is expanded, and the detection information increment between different sensors is further mined to achieve refined detection of small space targets.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for detecting small space targets by integrating microwave infrared information, characterized in that: The following steps are involved: S1, radar data preprocessing, radar target echo is processed by matched filtering and Doppler processing to generate a range-Doppler map, and the range-Doppler map is processed to form an RDPA map; S2, image segmentation algorithm based on threshold segmentation to extract and analyze target shape features of infrared images; S3. Design a spatial small target detection network that integrates RDPA image and infrared image information.

2. A method for detecting small space targets by integrating microwave infrared information according to claim 1, characterized in that: In S1, the range-Doppler map formula is: ； in represents convolution, Indicates the transmission signal. is the time reversal of the transmitted signal, represents the designed filter, K is the normalization factor, represents the output of the matched filter, Indicates the received echo signal, represents the first Pulses, t Indicates time, is the fast Fourier transform of the slow time dimension, represents the result of coherent accumulation, represents the Doppler frequency, represents the target speed, Indicates wavelength.

3. The method for detecting small space targets by integrating microwave infrared information according to claim 2, characterized in that: In S1, the range-Doppler image is processed as follows: in the range-Doppler image, the highest point of the peak is taken as the center point, and a square area around the center point is intercepted according to the side length d to form an RDPA image.

4. The method for detecting small space targets by integrating microwave infrared information according to claim 3 is characterized in that: In S2, the image segmentation algorithm specifically includes the following steps: S21, input image Convert to grayscale ; S22, determine the image extraction area threshold and segmentation threshold ; S23, grayscale image Convert to binary image ; ； S24, remove binary images Less than the extraction area threshold The connected domain of S25, mark greater than the segmentation threshold The connected domain of S26, return greater than the split threshold The starting position and row / column width of the connected component.

5. The method for detecting small space targets by integrating microwave infrared information according to claim 4 is characterized in that: In S2, when extracting shape features, the invariant moments, length and width in the shape features are selected as metrics, and the Hu invariant moment method is used to calculate 7 Hu invariant moments of the segmented area to construct a shape feature vector.

6. The method for detecting small space targets by integrating microwave infrared information according to claim 5, characterized in that: In S3, the spatial small target detection network includes an RDPA graph convolution module for extracting spectral features in the RDPA graph, an infrared convolution module for extracting background features in the infrared graph, and a fusion module for fusing the spectral features with the background features. The fusion module sends the fused joint features to the fully connected layer for classification; the RDPA graph is used as the input of the RDPA graph convolution module, and the infrared graph shape feature vector is used as the input of the infrared convolution module.

7. The method for detecting small space targets by integrating microwave infrared information according to claim 6 is characterized in that: In S3, the RDPA graph convolution module is provided with an attention module for weighting spectral features and a fusion module for fusing spectral features at different levels.

8. The method for detecting small space targets by integrating microwave infrared information according to claim 7, characterized in that: In S3, the infrared convolution module is provided with an attention module for weighting background features and a fusion module for fusing background features at different levels.

9. The method for detecting small space targets by integrating microwave infrared information according to claim 8, characterized in that: The formula of the attention module is: ； in, represents the input features, Represents the output attention mechanism features; Represents the convolution layer. The number of channels and dimensions of the feature remain unchanged after passing through this convolution layer. Responsible for generating attention weights at each feature point and comparing them with the input feature Multiply and exert attention impact.

10. The method for detecting small space targets by integrating microwave infrared information according to claim 9, characterized in that: The formula of the fusion module is: ； in, and represents the input of the fusion module, is the output of the fusion module, Describes the connection of the input feature map in the channel dimension, represents the batch normalization operation, Represents the activation function.

Images