Target matching method based on depth region perception and position regression

Abstract

The target matching method based on depth region perception and position regression is suitable for detection of an aerial carrier radar, a controller and a plurality of filters are installed on the aerial carrier, and the method comprises the steps that S101, echo signals preprocessed by the radar are acquired, and corresponding matched filters are designed according to waveforms emitted by the radar; s102, carrying out decomposition by the matched filter; s103, performing matched filtering processing on the decomposed matched filter to determine a matched filtering result; and S104, the target distance is quickly calculated according to the matched filtering result, and difference distance channel processing is added, so that the distance measurement precision of the target can be improved, and meanwhile, a relatively small calculation amount is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target matching, and in particular relates to a target matching method based on depth region perception and position regression. Background Art

[0002] Airborne synthetic aperture radar (SAR) imaging technology can be used as an auxiliary means to enhance the active offensive capability of carrier aircraft. For example, rapid strikes on specific ground targets require the use of SAR image matching technology to quickly find and identify targets; satellite navigation is susceptible to enemy interference during long-range operations, and SAR image matching technology is needed to achieve high-precision positioning information correction.

[0003] At present, target matching methods for engineering applications are mostly based on scale-invariant feature transformation methods. They achieve target matching by extracting key points from images and designing complex manual descriptors and matching mechanisms. However, there are common problems such as large amount of calculation, difficulty in achieving real-time matching, and coherent speckle noise and radiation intensity difference interfering with algorithm implementation. Summary of the invention

[0004] In view of this, the target matching method based on deep area perception and position regression of the present invention realizes end-to-end real-time target matching, and can find the precise position of the target image in the reference image while ensuring accuracy and speed.

[0005] A target matching method based on depth area perception and position regression is suitable for matching and positioning ground targets by airborne radar. The target matching method includes:

[0006] S101: Acquire a target image and a pre-stored reference image library, wherein the target image is a real-time image of a radar, and the reference image library includes reference images;

[0007] S102: extracting features of the target image and the reference image using a multi-scale feature pyramid network, and outputting a specific position of the target image in the reference image using a region perception and position regression model.

[0008] Beneficial Effects

[0009] By rotating the target image and adding noise to simulate the real scene and expand the training data, the preprocessed target image and reference image are input into the feature pyramid network to extract multi-scale features. The deep region perception module and position regression module are constructed through multi-layer convolutional network and fully connected network. The deep region perception module uses 1 / 8 scale image features to classify the position of the target image in the reference image and narrow the matching range. The position regression module uses 1 / 2 scale image features to perform secondary matching within the narrowed matching range and outputs the center position and offset of the target image in the reference image. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 It is a target matching flow chart of the depth region perception and position regression of the present invention. DETAILED DESCRIPTION

[0012] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0013] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0014] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0015] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure, and only show the components related to the present disclosure rather than being drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0016] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, one skilled in the art will appreciate that aspects may be practiced without these specific details.

[0017] like Figure 1 The target matching method based on deep area perception and position regression shown is suitable for matching and positioning of ground targets by airborne radar. The target matching method includes:

[0018] S101: Obtain a target image and a pre-stored reference image library, wherein the target image is a real-time image of a radar, and the reference image library includes reference images, for example, select a reference image in the SAR-AI Rcraft dataset according to matching requirements, crop the target from the reference image, generate the target image by rotation, translation, adding noise, etc., and create a training set and a test set;

[0019] S102: Use a multi-scale feature pyramid network to extract features of the target image and the reference image, and use the "region perception and position regression" model to output the specific position of the target image in the reference image. Generally, the "region perception and position regression" network model includes a region perception network model and a position regression network model. The purpose is to ensure the matching accuracy and robustness for rotation, translation, and noisy SAR images. Specifically:

[0020] S201: selecting a reference image from a reference image library according to a matching requirement in an open source dataset, such as a SAR-AIRcraft dataset;

[0021] S202: The target image and the reference image are respectively denoted as The corresponding image sizes are denoted as W t ×H t and W r ×H r ;

[0022] The multi-scale feature pyramid network is represented as f b (), feature extraction is performed on the target image and the reference image through a multi-scale feature pyramid network, and image features of 1 / 8 scale and 1 / 2 scale are obtained respectively, satisfying:

[0023] X t1 ,X t2 =f b (X t ),X r1 ,X r2 =f b (X r ) where X t1 is the 1 / 8 scale image feature of the target image, X t21 / 2 scale image features of the target image; X r1 is the 1 / 8 scale image feature of the reference image, X r2 1 / 2 scale image features of the reference image. 1 / 8 scale and 1 / 2 scale image features are used to distinguish objects of different scales and improve the versatility and accuracy of object positioning;

[0024] S103: 1 / 8 scale feature X of target image and reference image t1 and X r1 The 1 / 8 scale image features X of the target image are respectively input into the region-aware network model, and two separate convolutional networks in the region-aware network model are used to transform the 1 / 8 scale image features X of the target image into t1 and the 1 / 8 scale image feature X of the reference image r1 Adjust to the classification space and obtain the convolution feature and convolutional features Respectively expressed as:

[0025] and In the formula, Conv() is the convolution operation;

[0026] The cross-correlation feature is obtained by the "cross-correlation" method Wherein, * is a cross-correlation operation, for example, an operation similar to or the same as convolution;

[0027] Use the fully connected network model as a classifier to perform region classification and predict the region number where the target image is located in the reference image.

[0028] According to the area number Calculate the center point of the target image in the area where the reference image is located. The center point includes the horizontal and vertical coordinates (x c ,y c ) and width and height (w c ,h c );

[0029] S104: Obtain the corresponding regional feature map in the 1 / 2 scale feature of the region in the reference image according to the region perception network model And the image feature X t2 , Input the position regression network model, and use two separate convolutional networks in the position regression network model to transform the feature X t2 , Adjust to the regression space and get the regression features Regression Features

[0030]

[0031]

[0032] Among them, Conv() represents the convolution operation;

[0033] The regression cross-correlation feature is obtained through the "cross-correlation" method of the regression model.

[0034] Among them, * represents the cross-correlation operation;

[0035] Using the fully connected network model f r Perform position classification to obtain the specific position coordinates (x r ,y r ) and position coordinate offset (x rs ,y rs ):

[0036] In the formula, Represented as a fully connected network model f r The purpose of the calculated coordinate position and offset is: (x r ,y r ,x rs ,y rs ) can be directly obtained through the region perception network model and the position regression network model without extracting feature points. Therefore, the method of the present invention can perform image block matching without extracting feature points, can achieve real-time matching, and has high robustness to rotation and noise, and can achieve high-precision matching.

[0037] In the above, the number of channels of the target image and the reference image in S202 is 3, and preferably, both are color images.

[0038] As a specific implementation method provided in this case, it also includes optimizing the "regional perception and position regression" network model, which includes a region perception network model and a position regression network model, wherein:

[0039] The region-aware network model calculates the prediction loss L through the cross entropy loss function p , expressed as:

[0040] Where D train is the training sample set, y pi is the label of the region where the i-th target image is located in the region perception stage;

[0041] The position regression network model calculates the prediction loss L through the cross entropy loss function cls-x and the prediction loss L cls-y ,in,

[0042]

[0043] In the formula, (x ri ,y ri ) is the specific position coordinate of the i-th target image in the reference image, The position coordinates predicted by the model in the position regression stage.

[0044] Further, the optimization of the "regional awareness and location regression" network model includes:

[0045] The generalized intersection-over-union loss function is used to calculate the position offset prediction loss L reg , expressed as:

[0046] Where IoU is the ratio of the area of ​​the overlapped part of the target image real frame and the predicted frame to the total area of ​​the real frame and the predicted frame. It is 1 when they are completely overlapped and 0 when there is no overlap. Area represents the minimum area that can contain both the real frame and the predicted frame. U is the area of ​​the overlapped part of the target image real frame and the predicted frame.

[0047] Calculate the "matching loss sum" L of the "region awareness and location regression" network model, expressed as:

[0048] L=L p +L cls-x +L cls-y +L reg ;

[0049] The “matching loss and” L use the stochastic gradient descent method to optimize the network parameters of the “region perception and location regression” network model.

[0050] In general, the present invention can perform image block matching without extracting feature points through the design of multiple deep learning networks, can achieve real-time matching, and has high robustness to rotation and noise, can achieve high-precision matching. The method of the present invention is simple and effective, and can achieve homologous target matching accuracy within 3 pixels. It can be used for matching navigation of aircraft / unmanned aerial vehicles and matching positioning of time-sensitive targets on the ground.

[0051] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present disclosure should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A target matching method based on deep area perception and position regression, suitable for matching and positioning of ground targets by airborne radar, characterized in that: The target matching method comprises: S101: Acquire a target image and a pre-stored reference image library, wherein the target image is a real-time image of a radar, and the reference image library includes reference images; S102: extracting features of the target image and the reference image using a multi-scale feature pyramid network, and outputting a specific position of the target image in the reference image using a region perception and position regression model.

2. The target matching method according to claim 1, characterized in that: S102 includes: S201: Select a reference image from a reference image library in an open source dataset according to matching requirements; S202: The target image and the reference image are respectively recorded as The corresponding image sizes are denoted as W t ×H t and W r ×H r , R represents the real number space; The multi-scale feature pyramid network is represented as f b ( ), feature extraction is performed on the target image and the reference image through a multi-scale feature pyramid network, and image features of 1 / 8 scale and 1 / 2 scale are obtained respectively, satisfying: X t1 , X t2 =f b (X t ),X t1 , X t2 =f b (X r ), where X t1 is the 1 / 8 scale image feature of the target image, X t2 1 / 2 scale image features of the target image; X r1 is the 1 / 8 scale image feature of the reference image, X r2 1 / 2 scale image features of the reference image; S103: 1 / 8 scale feature X of the target image and the reference image t1 and X r1 The 1 / 8 scale image features X of the target image are respectively input into the region-aware network model, and two separate convolutional networks in the region-aware network model are used to transform the 1 / 8 scale image features X t1 and the 1 / 8 scale image feature X of the reference image r1 Adjust to the classification space and obtain the convolution feature and convolutional features Respectively expressed as: and Where Conv( ) is the convolution operation; The cross-correlation feature is obtained by cross-correlation method Where * is the cross-correlation operation; Use the fully connected network model as a classifier to perform region classification and predict the region number where the target image is located in the reference image. According to the area number Calculate the center point of the target image in the area where the reference image is located. The center point includes the horizontal and vertical coordinates (x c ,y c ) and width and height (w c ,h c ); S104: Obtaining, according to the region perception network model, a region feature map corresponding to the 1 / 2 scale feature of the region in the reference image And the image feature X t2 , Input the position regression network model, and use the two separate convolutional networks in the position regression network model to transform the feature X t2 , Adjust to the regression space and get the regression features Regression Features Among them, Conv( ) represents the convolution operation; The regression cross-correlation feature is obtained by the cross-correlation method of the regression model. Using the fully connected network model f r Perform position classification to obtain the specific position coordinates (x r ,y r ) and position coordinate offset (x rs ,y rs ): In the formula, Represented as a fully connected network model f r Calculated coordinate position and offset.

3. The target matching method according to claim 2, characterized in that: The number of channels of the target image and the reference image in S202 is 3.

4. The target matching method according to claim 3, characterized in that: The target image and the reference image in S202 are both color images.

5. The target matching method according to claim 4, characterized in that: It also includes optimizing the region perception and position regression network model, wherein the region perception and position regression network model includes a region perception network model, wherein: The region perception network model calculates the prediction loss L through the cross entropy loss function p , expressed as: Where D train is the training sample set, y pi is the label of the region where the i-th target image is located in the region perception stage.

6. The target matching method according to claim 5, characterized in that: The method also includes optimizing the region perception and position regression network model, wherein the region perception and position regression network model includes a position regression network model, and the position regression network model calculates the prediction loss L through a cross entropy loss function. cls-x and the prediction loss L cls-y ,in, In the formula, (x ri ,y ri ) is the specific position coordinate of the i-th target image in the reference image, The position coordinates predicted by the model in the position regression stage.

7. The target matching method according to claim 6, characterized in that: Optimizing the region perception and location regression network model includes: The generalized intersection-over-union loss function is used to calculate the position offset prediction loss L reg , expressed as: Where IoU is the ratio of the area of ​​the overlapped part of the target image real frame and the predicted frame to the total area of ​​the real frame and the predicted frame. It is 1 when they are completely overlapped and 0 when there is no overlap. Area represents the minimum area that can contain both the real frame and the predicted frame. U is the area of ​​the overlapped part of the target image real frame and the predicted frame. Calculate the matching loss and L of the region perception and location regression network models, expressed as: L=L p +L cls-x +L cls-y +L reg ; The matching loss and L use the stochastic gradient descent method to optimize the network parameters of the region perception and location regression network models.

Images