A video SAR anti-interference imaging method suitable for light and small platforms

Abstract

The present invention discloses a video SAR anti-interference imaging method suitable for light and small platforms, and belongs to the field of radar imaging processing. In a complex electromagnetic interference environment, the method of the present invention realizes the interlaced extraction of the projection grid positions of continuous frames by constructing time-space sampling particles to obtain a small-size image with full aperture resolution. K consecutive frames of small-size images that may be interfered with are input into an anti-interference fusion reconstruction network, and down-sampled through multiple convolutional layers to extract the features of the K frame images. The extracted feature map is then input into the anti-interference network module to remove interference. After upsampling processing through multiple convolutional layers, an anti-interference fusion SAR image is finally reconstructed. The method of the present invention can provide high-resolution SAR imaging services for light and small platforms under electromagnetic interference conditions, significantly improving the anti-interference capability and imaging quality.

Description

Technical Field

[0001] The present invention relates to the field of radar imaging technology, and in particular to a video SAR (synthetic aperture radar) anti-interference imaging method suitable for a light and small platform. Background Art

[0002] Lightweight and small UAVs are becoming increasingly important as resource-constrained platforms in areas such as environmental remote sensing and disaster prevention. Video synthetic aperture radars embedded in these UAVs have enabled all-weather, high-resolution microwave imaging, significantly advancing the state of the art. However, resource constraints have resulted in low navigation accuracy, insufficient computing power, and limited communication bandwidth, which poses challenges to high-precision imaging.

[0003] The multi-frame data output by video synthetic aperture radar shows spatiotemporal correlation. A single frame image has a short imaging aperture, while adjacent multi-frame images are combined into a longer imaging aperture, providing higher azimuth resolution. In the face of complex electromagnetic interference environments, interference signals may affect imaging quality. Therefore, it is crucial to develop effective anti-interference technology. Optimizing the spatiotemporal sampling algorithm and improving anti-interference capabilities will provide new solutions for video synthetic aperture radar applications on resource-constrained platforms. Summary of the invention

[0004] In order to overcome the resource limitation of existing light and small-sized mounted video SAR platforms and improve the imaging quality in electromagnetic interference environments, the present invention provides a video SAR anti-interference imaging method suitable for light and small-sized platforms.

[0005] A video SAR anti-interference imaging method suitable for a light and small platform, comprising:

[0006] Step 1, defining the sampling position (also called sampling particle position) and sampling density (also called sampling particle density) of the time sampling particles in the full projection aperture;

[0007] Based on the sampling position and sampling density, a sparse sampling point set {(t i ,s j )}, where t i and j They represent the sampling positions in the time and space dimensions respectively, and the subscripts i and j represent the sampling position indexes in the time and space dimensions respectively;

[0008] Step 2, constructing an image fusion network including multiple downsampling layers, wherein each downsampling layer sequentially includes several convolution blocks and a maximum pooling layer, and each convolution block sequentially includes a convolution layer, a batch normalization layer, and an activation layer;

[0009] Step 3, constructing an anti-interference network for feature map denoising, which is used to perform channel-by-channel denoising on the output feature map of the last downsampling layer of the image fusion network, and output the denoised feature map;

[0010] Step 4, constructing an image reconstruction network including multiple upsampling layers and a reconstruction output layer, wherein each upsampling layer includes a plurality of convolution blocks and an upconvolution layer in sequence; the reconstruction output layer is used to reconstruct the multi-channel feature map into a single-channel fused spatiotemporal interleaved sampling synthetic aperture radar image;

[0011] Step 5, setting the convolution feature map of the downsampling layer of the image fusion network to be consistent with the output feature map of the corresponding upsampling layer in the image reconstruction network in the spatial dimension, the convolution feature map of the downsampling layer and the output feature map of the upsampling layer form a one-to-one correspondence, and then the convolution feature map of the corresponding downsampling layer and the output feature map of the upsampling layer are spliced ​​in the channel dimension to perform a jump connection between the upsampling layer and the downsampling layer;

[0012] Step 6, constructing an anti-interference fusion reconstruction network based on the image fusion network, the anti-interference network and the image reconstruction network, inputting K frames of spatiotemporal staggered sampling synthetic aperture radar SAR images into the image fusion network, obtaining an output feature map of the image fusion network after multiple downsampling layers, and inputting it into the anti-interference network for denoising, and then inputting it into the image reconstruction network, obtaining a reconstructed K-frame anti-interference fusion image after skip connection, multiple upsampling layers and reconstruction output layer; wherein each frame of spatiotemporal staggered sampling SAR image corresponds to a sampling point in the sparse sampling point set, and K is a preset value, which is greater than or equal to 2;

[0013] Step 7, initialize the anti-interference fusion reconstruction network and perform network parameter training. When the set training convergence conditions are met (such as the number of training times reaches a preset upper limit, or the value of the set loss function converges), the trained anti-interference fusion reconstruction network is obtained;

[0014] Step 8, input the K-frame spatiotemporal staggered sampling SAR images into the trained anti-interference fusion reconstruction network, obtain the reconstructed K-frame anti-interference fusion images based on its output, and adjust the sampling position and sampling density of the time sampling particles based on the SAR image quality evaluation index feedback of the reconstructed K-frame anti-interference fusion images.

[0015] Furthermore, in step 1, the sampling density of the time dimension is The sampling density of the spatial dimension is N t 、N s Respectively represent the number of samples in time and space dimensions, T t , T s are the sampling unit time of the time dimension and the sampling unit space of the space dimension respectively.

[0016] Furthermore, in step 2, the activation layer of the convolution block adopts the Leaky ReLU activation function.

[0017] Furthermore, in step 2, the number of output channels of the downsampling layer increases layer by layer.

[0018] Furthermore, in step 2, the convolution kernel size of the convolution block of each downsampling layer is set to 3×3, the padding is set to 1, and the maximum pooling layer step size of each downsampling layer is set to 2.

[0019] Furthermore, in step 3, the anti-interference network includes: a convolutional pooling structure in which several convolutional blocks and maximum pooling layers are alternated in sequence, a reshaping layer, at least two fully connected layers and a denoising layer, wherein the convolutional pooling structure is used to perform a channel-by-channel convolutional pooling operation on each channel of the output feature map of the last downsampling layer of the image fusion network; the reshaping layer is used to adjust the output feature map of the convolutional pooling structure to the target size before sending it to the fully connected layer, and the last fully connected layer uses a Sigmoid activation function to map its input feature map to a probability value between 0 and 1. To characterize whether the feature map is noise, the denoising layer is based on the probability value The relationship between the size of the preset threshold τ and the output feature map f of the last fully connected layer c Perform denoising: If Then for the output feature map f c Use zero filling or specified value replacement to obtain the effective feature map of the current channel; otherwise, directly replace f c is regarded as the valid feature map of the current channel, where the subscript c is the channel index of the feature map processed by the anti-interference network; finally, the output feature map of the anti-interference network is obtained based on the valid feature maps of all channels.

[0020] Furthermore, in step 3, the threshold τ is set to τ = 0.5 ± Δ τ , Δ τ Indicates the allowable deviation value.

[0021] Furthermore, in step 4, for each upsampling layer, the step size of its upper convolution layer is defined as d, the resolution of the upsampled feature map is d times the resolution of the input feature map of the upsampling layer, and the number of output channels of the upsampling layer is half the number of input channels.

[0022] Furthermore, in step 4, the reconstructed output layer sequentially includes a plurality of convolution blocks and a convolution layer with a convolution kernel of 1×1 and a Softmax function.

[0023] Furthermore, the number of downsampling layers of the image fusion network is the same as the number of upsampling layers of the image reconstruction network.

[0024] Furthermore, in step 5, the jump connection between the downsampling layer and the upsampling layer is set to:

[0025] According to the direction of forward propagation, the downsampling layers of the image fusion network are defined as the 1st to the Nth layers in sequence, and the upsampling layers of the image reconstruction network are defined as the 1st to the Nth layers in sequence;

[0026] For any i-th downsampling layer, a skip connection is set between the output feature map of the last convolution block of downsampling layer i and the output feature map of the upconvolution layer of the N-1-i-th upsampling layer.

[0027] Further, step 7 includes:

[0028] Step 701, using Gaussian distribution to randomly generate network parameters of the anti-interference fusion reconstruction network, including weight W and bias b;

[0029] Step 702: input each set of training sample images into the anti-interference fusion reconstruction network, and output the reconstructed anti-interference fusion image through forward propagation, which is recorded as the output prediction image.

[0030] Step 703: Output the predicted image The difference between the real label image y is quantified by the loss function;

[0031] Step 704, using a back propagation algorithm to calculate the gradient of the loss function L with respect to the network parameters, and then using the chain rule to propagate the gradient backward layer by layer, and using an optimization algorithm to update the weight W and bias b based on the calculated gradient;

[0032] Step 705, determine whether the number of iterative updates reaches the preset iteration upper limit; if so, end the sequence and obtain a trained anti-interference fusion reconstruction network based on the currently updated weights W and bias b; otherwise, return to execute step 702 based on the currently updated weights W and bias b.

[0033] Furthermore, in step 8, SAR image quality evaluation indicators include but are not limited to: signal-to-noise ratio (SNR), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), etc.

[0034] Further, in step 8, adjusting the sampling position and sampling density of the time sampling particles based on the SAR image quality evaluation index feedback includes:

[0035] The quality deviation metric between the reconstructed anti-interference fusion image and the true label image is evaluated based on multiple SAR image quality evaluation indicators, and the quality score of each reconstructed anti-interference fusion image is obtained based on the weighted sum of all quality deviation metric values;

[0036] The update amount ΔP of the sampling position is obtained by multiplying the gradient of the current sampling position P by the sampling position learning rate (preset value) according to the mean value of the quality score of the anti-interference fusion image reconstructed by K frames;

[0037] The sampling density update amount ΔD is obtained by multiplying the gradient of the current sampling density D and the sampling density learning rate (preset value) by the mean of the quality scores of the anti-interference fusion images reconstructed by the K frames;

[0038] According to P'=P+ΔP and D'=D+ΔD, the updated sampling position P' and the updated sampling density D' are obtained.

[0039] Then, based on the updated sampling position and sampling density, resample in the full projection aperture to obtain a new sparse sampling point set, and then use the spatiotemporal interleaved projection synthetic aperture radar image corresponding to the K consecutive sparse sampling points in the new sparse sampling point set as the new input image of the anti-interference fusion reconstruction network, and continue to execute step 8.

[0040] The technical solution provided by the present invention brings at least the following beneficial effects:

[0041] The present invention provides a video SAR anti-interference imaging method suitable for a light and small platform, which reduces the calculation amount of the platform and enhances the anti-interference ability of imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 A flowchart of a video SAR anti-interference imaging method suitable for a light and small platform provided in an embodiment of the present invention.

[0044] Figure 2 A schematic diagram of a network structure used in an embodiment of the present invention.

[0045] Figure 3 The following are comparison diagrams of video SAR imaging before and after processing by the method proposed in the embodiment of the present invention, wherein (3-a) is the original SAR image subjected to electromagnetic interference, and (3-b) is the anti-interference fusion reconstructed SAR image. DETAILED DESCRIPTION

[0046] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.

[0047] In one embodiment, Figure 1 As shown, a video SAR anti-interference imaging method suitable for a light and small platform provided in an embodiment of the present invention is implemented by the following steps:

[0048] S1. Constructing spatiotemporal sampling particles, that is, constructing a full projection aperture spatiotemporal staggered sampling algorithm, by defining the position and sampling density of spatiotemporal sampling particles in the full projection aperture to form a sparse sampling point set. The position of the spatiotemporal sampling particle is (t i ,s j ), where t i and j The subscripts i and j represent the sampling positions in the time and space dimensions respectively. The sampling density is also set and To control the number of sampled particles per unit time or space. t 、N s Respectively represent the number of samples in time and space dimensions, T t , T s The sampling units of time dimension and space dimension are respectively the sampling units of time dimension and space dimension. Finally, the sparse sampling point set S is composed of S = {(t i ,s j )|t i ∈[0,T t ],s j ∈[0,T s ]}, thereby achieving efficient sampling of spatiotemporal data.

[0049] In this embodiment, a four-frame uniform sampling algorithm is used to divide the full projection aperture into a uniform space-time network, and a sampling particle is defined at each network node, thereby obtaining four sparse sampling points, corresponding to the sampling positions in the time and space dimensions, namely (t1, s1), (t2, s2), (t3, s3), (t4, s4), where t i Indicates the sampling position in the time dimension, s i Represents the sampling position in the spatial dimension, achieving uniform coverage across the full aperture.

[0050] S2. Construct an image fusion network, which includes several downsampling layers, each of which includes M convolution blocks and a maximum pooling layer, where M≥2, and each convolution block includes a convolution layer, a batch normalization layer, and an activation function layer in sequence. In this embodiment, the activation function layer uses a ReLus activation function, and the convolution kernel size of the convolution layer is set to K×K, where K is a preset value, and each downsampling layer includes two convolution blocks. The weight parameters of the image fusion network are continuously updated with training to improve the feature fusion capability.

[0051] In this embodiment, the image fusion network input is composed of 4 frames of spatiotemporal interleaved projection synthetic aperture radar images obtained in S1, which are passed through 4 downsampling layers. The downsampling layer consists of two modules, each of which contains 2 3×3 convolution kernels and ReLus. Then, after a 2×2 convolution layer and a maximum pooling operation with a step size of 2, the resolution of the feature map after the downsampling layer is reduced to half of the original resolution, and the number of output feature channels is set to twice the number of input feature channels. After four downsampling layers, four sets of feature maps of different sizes are obtained. In this embodiment, in order to ensure that the feature maps before and after the convolution layer have the same resolution, the padding calculation is 1.

[0052] S3. Construct an anti-interference network, which includes several convolution blocks, maximum pooling layers and fully connected layers. Among them, the convolution kernel size of the convolution block is expressed as d×d, the step size of the maximum pooling layer is set to 2, and the activation function used by each convolution block is set to the Leaky ReLU activation function. In the embodiment of the present invention, in the anti-interference network, the output feature map of the last downsampling layer is processed channel by channel to obtain the effective feature map of each channel, and then the output feature map of the anti-interference network is obtained based on the effective feature maps of all channels. In the anti-interference network, the input feature map is first convoluted and pooled by alternating convolution blocks and maximum pooling layers, and then the feature vector output by the last maximum pooling layer is reshaped by a reshaping layer (Resize layer). After adjusting to the feature map of the target size, it is sent to the fully connected layer. The last fully connected layer uses the Sigmoid activation function to map its input feature map to a probability value between 0-1. To characterize whether the feature map is noise, the denoising layer (the output layer of the anti-interference network) is based on the probability value The relationship between the size of the preset threshold τ and the corresponding denoising process is performed on the output feature map of the last fully connected layer: If the output feature map is judged to be noisy, it is considered that the output feature map has noise, otherwise it is considered to be a valid feature map of the current channel; for the output feature map judged to be noisy, zero padding or replacement is used to obtain the valid feature map of the current channel; finally, the output feature map of the anti-interference network is obtained based on the valid feature maps of all channels. Among them, the weight parameters of the anti-interference network are continuously updated with training to improve the ability to distinguish noise.

[0053] In this embodiment, an anti-interference network is constructed, which includes 2 convolutional layers, 2 maximum pooling layers and 2 fully connected layers, where the size of the convolution kernel is 3×3 and the step size of the maximum pooling layer is 2. Each convolutional layer is followed by a Leaky ReLU activation function (the parameter is set to α=0.2), and then the feature map is compressed into a feature vector through a global average pooling layer and a Resize layer. Then, the feature vector is mapped through two fully connected layers, and finally a Sigmoid activation function is used to output a probability value between 0 and 1. Set the threshold τ = 0.5, if If the feature map is judged as noise, it is considered to have noise, otherwise it is considered to be a valid feature map. For the feature map judged as noise, it is filled with zeros to achieve the denoising function.

[0054] S4. Construct an image reconstruction network, which includes multiple upsampling layers and a reconstruction output layer, wherein each upsampling layer includes multiple convolution blocks and an up-convolution layer (Up-conv, also called deconvolution layer) in sequence, and the reconstruction output layer is used to reconstruct the multi-channel feature map into a single-channel fused spatiotemporal interleaved sampling synthetic aperture radar image.

[0055] In this embodiment, the reconstructed output layer sequentially includes a plurality of convolution blocks and a convolution layer with a 1×1 convolution kernel and a Softmax function.

[0056] The resolution of the feature map after upsampling in each upsampling layer is d times the original resolution, and the number of output channels after upsampling is half the number of input channels, where d is a preset value corresponding to the step size of the up-convolution layer. In this embodiment, the convolution kernel size of the Up-conv layer is set to 2x2, and the step size is set to 2. After the four upsampling layers, two convolution blocks, a 1x1 convolution layer, and a Softmax function are also set to realize the reconstruction of the multi-channel feature map into a single-channel fused spatiotemporal interleaved sampling synthetic aperture radar image.

[0057] S5. Construct a jump connection structure between the image fusion network and the image reconstruction network.

[0058] When constructing the image fusion network and the image reconstruction network, the feature maps of the corresponding layers have the same size in spatial dimension. The feature map of a layer in the image fusion network is And the feature map of the corresponding layer in the image reconstruction network is Where H and W represent the height and width of the feature map, respectively, and C f and C r is the number of channels of the corresponding layer, Represents the real number domain. By performing jump connections through splicing, a new feature map is obtained The concatenation operation can be expressed as:

[0059] F combined =concat(F fusion ,F recon ,axis=-1)

[0060] Here, axis represents the specified dimension in the feature map, that is, axis=-1 means concatenating two feature maps in the channel dimension.

[0061] In this embodiment, the feature map obtained after four downsampling layers in step S2 and the feature map obtained after four upsampling layers in step S4 have the same size in the corresponding number of layers, and can be skip-connected by channel splicing. That is, the output feature map of the last convolution block of the downsampling layer is skip-connected with the output feature map of the upper convolution layer of the corresponding upsampling layer by channel splicing.

[0062] S6, build an anti-interference fusion reconstruction network, that is, input K frames of spatiotemporal staggered sampling synthetic aperture radar graphics into the image fusion network. The input image is Where H and W are the height and width of the image, respectively, and C is the number of channels of the image. The feature map obtained after the i-th downsampling layer is It can be expressed as: F i = Downsample(F i-1 ), i=1,2,...,N,C i , H i , W i are the number of channels, height and width of the image in the i-th downsampling layer respectively. N The input is sent to the anti-interference network and the Sigmoid activation function is used to determine whether the feature map is a noise image. The output result is:

[0063]

[0064] Among them, σ() represents the Sigmoid function, W F is the weight and b is the bias.

[0065] Fill or replace the feature map identified as noise with zeros, that is:

[0066]

[0067] Finally, Foutput The input image reconstruction network consists of N upsampling layers. Through skip connections, the feature map from the image fusion network is concatenated with the reconstructed feature map of the corresponding layer. The output feature map of the upsampling layer is The calculation process is:

[0068]

[0069] in, Represent the output feature maps of the upsampling layers of the i-th layer and the i-1-th layer, respectively. Then it corresponds to the output feature map of the anti-interference network. The above formula shows that the feature map is first upsampled to obtain the feature map F recon , and then the convolution feature map F of the corresponding downsampling layer fusion With F recon Concatenate by channel dimension.

[0070] Then, based on the output feature map of the reconstruction network, the final reconstructed anti-interference fusion SAR image I is obtained. output , whose image dimension is consistent with the input image.

[0071] S7, training the anti-interference fusion reconstruction network, initializing the network parameters, training the constructed anti-interference fusion reconstruction network, and obtaining the weight parameters of the anti-interference fusion reconstruction network;

[0072] S8. Reconstruct the spatiotemporal sampling particles according to the image quality evaluation index feedback, input the K frames of spatiotemporal interleaved sampling SAR images into the combined network with the weight parameters loaded, obtain the interference-resistant synthetic aperture radar image, and use the SAR image quality evaluation index to feedback and adjust the position and sampling density of the spatiotemporal sampling particles in the spatiotemporal sampling algorithm.

[0073] In one embodiment, the anti-interference fusion reconstruction network constructed by the embodiment of the present invention is as follows: Figure 2As shown, it includes 4 downsampling layers, each of which includes two convolution blocks and a maximum pooling layer, and the output of the last downsampling layer is input into the anti-interference network after a convolution block; that is, the image fusion network in this embodiment includes: convolution block 1, convolution block 2, maximum pooling layer 1, convolution block 3, convolution block 4, maximum pooling layer 2, convolution block 5, convolution block 6, maximum pooling layer 3, convolution block 7, convolution block 8 and maximum pooling layer 4 in sequence; wherein, the number of output channels of convolution block 1 and convolution block 2 is set to 32, the number of output channels of convolution block 3 and convolution block 4 is set to 64, the number of output channels of convolution block 5 and convolution block 6 is set to 128, and the number of output channels of convolution block 7 and convolution block 8 is set to 256, i.e., it increases layer by layer. The anti-interference network includes: convolution block 9, maximum pooling layer 5, convolution block 10, maximum pooling layer 6, reshaping layer (performing Resize operation), fully connected layer 1, fully connected layer 2 with Sigmoid activation function and denoising layer; wherein, the number of output channels of convolution block 9 is set to 64, and the number of output channels of convolution block 10 is set to 128; the Resize operation is used to reshape the output feature map of the maximum pooling layer 6 into a feature vector of the target dimension and then input it into the fully connected layer 1; the image reconstruction network includes 4 upsampling layers and two convolution blocks and a convolution layer with a softmax function (the convolution kernel is 1×1), that is, the image reconstruction network in this embodiment The network includes: convolution block 11, convolution block 12, upper convolution layer 1, convolution block 13, convolution block 14, upper convolution layer 2, convolution block 15, convolution block 16, upper convolution layer 3, convolution block 17, convolution block 18, upper convolution layer 4, convolution block 19, convolution block 20 and a convolution layer with a softmax function; the number of output channels of convolution blocks 11 and 12 is set to 512, the number of output channels of convolution blocks 13 and 14 is set to 256, the number of output channels of convolution blocks 15 and 16 is set to 128, the number of output channels of convolution blocks 17 and 18 is set to 64, and the number of output channels of convolution blocks 19 and 20 is set to 32.

[0074] The jump structure between the image fusion network and the image reconstruction network is as follows: the output feature map of convolution block 2 and the upper convolution layer 4 is concatenated according to the channel dimension and used as the input feature map of convolution block 19; the output feature map of convolution block 4 and the upper convolution layer 3 is concatenated according to the channel dimension and used as the input feature map of convolution block 17, that is, the input feature map of the last upsampling layer; the output feature map of convolution block 6 and the upper convolution layer 2 is concatenated according to the channel dimension and used as the input feature map of convolution block 15, that is, the input feature map of the second to last upsampling layer; the output feature map of convolution block 8 and the upper convolution layer 1 is concatenated according to the channel dimension and used as the input feature map of convolution block 13, that is, the input feature map of the first upsampling layer.

[0075] See also Figure 2In the anti-interference fusion reconstruction network shown, in this embodiment, the 4 frames of spatiotemporal uniformly sampled SAR images obtained in step S1 are input into the image fusion network, and the image size is 1×400×400 (SAR image is a single-channel image). After the four downsampling layers of the image fusion network, a set of feature maps with sizes of 32×200×200, 64×100×100, 128×50×50 and 256×25×25 are obtained respectively, and the feature map of the last layer of 256×25×25 is sent to the anti-interference network to perform channel-by-channel denoising. In the anti-interference network, the Sigmoid activation function is used to determine whether the feature map is a noise image, and the noise image is filled with zeros. At this time, the size and number of channels of the feature map do not change after passing through the anti-interference network. The denoised 256×25×25 feature map is input into the image reconstruction network. The feature map obtained after four upsampling layers is jump-connected with the corresponding downsampling layers. Finally, the multi-channel feature map is reconstructed into a 1x400x400 anti-interference fused SAR image through a 1x1 convolution operation and softmax.

[0076] In one embodiment, step S7 specifically includes:

[0077] S71. After constructing the anti-interference fusion reconstruction network, randomly generate initial network weights W and bias b. Preferably, Gaussian distribution is used to initialize the weight parameters, i.e., W i，j ~N(0, 0.012), the bias b is initialized to zero to simplify the network training. i,j represents the weight of the jth neuron in the i-th layer of the network.

[0078] S72. For each set of training sample images (x, y), where x represents the input image and y represents the corresponding true label image, it is first forward propagated through each layer of the network. After a series of convolutional layers, activation layers, and pooling layers, the predicted anti-interference image is finally output. Right now Where f represents the forward propagation function of the network.

[0079] S73. Output image The difference between the true label image y is quantified by the loss function L.

[0080] Preferably, a binary cross-entropy loss function is used, and its formula is:

[0081]

[0082] Where N' represents the number of sample images.

[0083] S74. Use the back propagation algorithm to calculate the gradient of the loss function L with respect to each network parameter.

[0084] First, the gradient of the output layer is calculated, and then the error is propagated backward layer by layer through the chain rule. Based on the calculated gradient, the weights and biases are updated using an optimization algorithm.

[0085] S75, determining whether the number of iterative updates reaches a preset number N" times;

[0086] S76. If the iteration reaches N” times, save the currently updated network weight W and bias b;

[0087] S77. If the number of iterations does not reach N", the process returns to step S72.

[0088] In this embodiment, the back propagation algorithm is a method used in combination with an optimization method to train a deep learning network. The method calculates the gradient of the loss function for all weights in the network, and this gradient is fed back to the optimization method to update the weights to minimize the loss function.

[0089] In one embodiment, step S8 specifically includes:

[0090] S81, input the K frames of spatiotemporal staggered sampling SAR images into the anti-interference fusion reconstruction network loaded with the network weights W and bias b obtained after training in step 7, and obtain the output image

[0091] S82. Using SAR image quality evaluation indicators to feedback output images The quality of the image is measured by the signal-to-noise ratio (SNR), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), etc.

[0092] S83. According to the feedback of the quality evaluation index, adjust the position and sampling density of the spatiotemporal sampling particles in the spatiotemporal sampling algorithm. Set the new sampling particle position P' and sampling density D', and update the sampling parameters by optimizing the objective function:

[0093] P'=P+ΔP、D'=D+ΔD

[0094] Among them, ΔP and ΔD are the update amounts calculated based on the quality evaluation index.

[0095] S84. Using the updated spatiotemporal sampling particles P' and sampling density D', redefine the positions and sampling density of the spatiotemporal sampling particles in the full projection aperture to achieve a new sparse sampling point set:

[0096] S'={(t i ',s' j )|t i'∈P',s' j ∈D'}

[0097] The process returns to S81 and iterates until an output image of ideal quality is obtained. This image is the interference-resistant fusion reconstructed image.

[0098] In this embodiment, three quality evaluation indicators, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and equivalent number of views (ENL), are selected to evaluate the output image from three aspects: overall image quality, image structural similarity and signal-to-noise ratio. According to the feedback of the quality evaluation index, the position and sampling density of the spatiotemporal sampling particles in the four-frame uniform spatiotemporal sampling algorithm in step S1 are adjusted, the new sampling particle positions P' and D' are set, and the sampling parameters are updated by optimizing the objective function:

[0099]

[0100] Among them, Y represents the noise-free contrast image, Q represents the output image The quality score of the product is w1, w2, and w3 are the weights of each indicator, satisfying w1+w2+w3=1.

[0101]

[0102] Among them, ΔP and ΔD are the update amounts calculated based on the quality evaluation index, α and β represent the learning rate or step size, It represents the gradient of the quality score Q with respect to the sampling particle position P and the sampling particle density D.

[0103] The updated spatiotemporal sampling particles P' and sampling density D' are used to realize a new sparse sampling point set. The process returns to step S81 for loop iteration.

[0104] After completing the above process, we can get Figure 3 The anti-interference fusion reconstructed SAR image shown in (3-b) in FIG. 1 is shown in FIG. 1 , where (3-a) is the original SAR image subjected to electromagnetic interference. Figure 3 It can be seen from the comparison diagram that the anti-interference capability and imaging quality of the anti-interference fused SAR image reconstructed by the embodiment of the present invention are significantly improved.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0106] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A video SAR anti-interference imaging method suitable for light and small platforms, characterized in that: The following steps are involved: Step 1, defining the sampling position and sampling density of the spatiotemporal sampling particles in the full projection aperture; Based on the sampling position and sampling density, a sparse sampling point set {(t i ,s j )}, where t i and j They represent the sampling positions in the time and space dimensions respectively, and the subscripts i and j represent the sampling position indexes in the time and space dimensions respectively; Step 2, constructing an image fusion network including multiple downsampling layers, wherein each downsampling layer sequentially includes several convolution blocks and a maximum pooling layer, and each convolution block sequentially includes a convolution layer, a batch normalization layer, and an activation layer; Step 3, constructing an anti-interference network for feature map denoising, which is used to perform channel-by-channel denoising on the output feature map of the last downsampling layer of the image fusion network, and output the denoised feature map; Step 4, constructing an image reconstruction network including multiple upsampling layers and a reconstruction output layer, wherein each upsampling layer includes a plurality of convolution blocks and an upconvolution layer in sequence; the reconstruction output layer is used to reconstruct the multi-channel feature map into a single-channel fused spatiotemporal interleaved sampling synthetic aperture radar image; Step 5, setting the convolution feature map of the downsampling layer of the image fusion network to be consistent with the output feature map of the corresponding upsampling layer in the image reconstruction network in the spatial dimension, the convolution feature map of the downsampling layer and the output feature map of the upsampling layer form a one-to-one correspondence, and then the convolution feature map of the corresponding downsampling layer and the output feature map of the upsampling layer are spliced ​​in the channel dimension to perform a jump connection between the upsampling layer and the downsampling layer; Step 6, constructing an anti-interference fusion reconstruction network based on the image fusion network, the anti-interference network and the image reconstruction network, inputting K frames of spatiotemporal staggered sampling synthetic aperture radar SAR images into the image fusion network, obtaining an output feature map of the image fusion network after multiple downsampling layers, and inputting it into the anti-interference network for denoising, and then inputting it into the image reconstruction network, obtaining a reconstructed K-frame anti-interference fusion image after skip connection, multiple upsampling layers and reconstruction output layer; wherein each frame of spatiotemporal staggered sampling SAR image corresponds to a sampling point in the sparse sampling point set, and K is a preset value, which is greater than or equal to 2; Step 7, initialize the anti-interference fusion reconstruction network and perform network parameter training. When the set training convergence conditions are met, a trained anti-interference fusion reconstruction network is obtained; Step 8, input the K-frame spatiotemporal interleaved sampling SAR images into the trained anti-interference fusion reconstruction network, obtain the reconstructed K-frame anti-interference fusion images based on its output, and adjust the sampling position and sampling density of the spatiotemporal sampling particles based on the SAR image quality evaluation index feedback of the reconstructed K-frame anti-interference fusion images.

2. A video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 1, the sampling density of the time dimension is The sampling density of the spatial dimension is N t 、N s Respectively represent the number of samples in time and space dimensions, T t , T s are the sampling unit time of the time dimension and the sampling unit space of the space dimension respectively.

3. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 2, the convolution kernel size of the convolution block of each downsampling layer is set to 3×3, the padding is set to 1, and the maximum pooling layer stride of each downsampling layer is set to 2.

4. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 3, the anti-interference network includes: a convolutional pooling structure in which several convolutional blocks and maximum pooling layers are alternated in sequence, a reshaping layer, at least two fully connected layers and a denoising layer, wherein the convolutional pooling structure is used to perform channel-by-channel convolutional pooling operations on each channel of the output feature map of the last downsampling layer of the image fusion network; the reshaping layer is used to adjust the output feature map of the convolutional pooling structure to the target size before sending it to the fully connected layer, and the last fully connected layer uses the Sigmoid activation function to map its input feature map to a probability value between 0 and 1 To characterize whether the feature map is noise, the denoising layer is based on the probability value The relationship between the size of the preset threshold τ and the output feature map f of the last fully connected layer c Perform denoising: If Then for the output feature map f c Use zero filling or specified value replacement to obtain the effective feature map of the current channel; otherwise, directly replace f c is regarded as the valid feature map of the current channel, where the subscript c is the channel index of the feature map processed by the anti-interference network; finally, the output feature map of the anti-interference network is obtained based on the valid feature maps of all channels.

5. A video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 4, characterized in that: In step 3, the threshold τ is set to τ = 0.5 ± Δ τ , Δ τ Indicates the allowable deviation value.

6. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 4, for each upsampling layer, the step size of its upper convolution layer is defined as d, the resolution of the upsampled feature map is d times the resolution of the input feature map of the upsampling layer, and the number of output channels of the upsampling layer is half the number of input channels.

7. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 4, the reconstructed output layer sequentially includes multiple convolution blocks and a convolution layer with a 1×1 convolution kernel and a Softmax function.

8. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 5, the skip connection between the downsampling layer and the upsampling layer is set as: The number of downsampling layers of the image fusion network is the same as the number of upsampling layers of the image reconstruction network. According to the direction of forward propagation, the downsampling layers of the image fusion network are defined as the 1st to the Nth layers in sequence, and the upsampling layers of the image reconstruction network are defined as the 1st to the Nth layers in sequence; For any i-th downsampling layer, a skip connection is set between the output feature map of the last convolution block of downsampling layer i and the output feature map of the upconvolution layer of the N-1-i-th upsampling layer.

9. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 8, the SAR image quality evaluation indicators include: signal-to-noise ratio, peak signal-to-noise ratio and structural similarity index.

10. The video SAR anti-interference imaging method suitable for a light and small platform as claimed in claim 1, characterized in that: In step 8, adjusting the sampling position and sampling density of the spatiotemporal sampling particles based on the feedback of the SAR image quality evaluation index includes: The quality deviation metric between the reconstructed anti-interference fusion image and the true label image is evaluated based on multiple SAR image quality evaluation indicators, and the quality score of each reconstructed anti-interference fusion image is obtained based on the weighted sum of all quality deviation metric values; The update amount ΔP of the sampling position is obtained by multiplying the gradient of the current sampling position P by the mean value of the quality score of the anti-interference fusion image reconstructed by K frames and the learning rate of the sampling position; The sampling density update amount ΔD is obtained by multiplying the gradient of the current sampling density D and the sampling density learning rate by the mean of the quality scores of the anti-interference fusion images reconstructed by K frames; According to P'=P+ΔP and D'=D+ΔD, the updated sampling position P' and the updated sampling density D' are obtained.

Images