A cnn denoising method based on range-doppler information

Abstract

The application discloses a CNN denoising method based on distance-Doppler information, comprising the following steps: acquiring continuous multiple frames of radar distance-Doppler images, and constructing a convolutional neural network model; introducing an effective loss solving function to the convolutional neural network model for training; inputting a current frame image, a previous frame image of the current frame and a previous two frame image of the current frame into the trained convolutional neural network model to obtain a denoised image of the current frame image, wherein the convolutional neural network model comprises a self-adaptive moving encoder module combined with a convolution block attention module, effectively avoiding the problems of redundant parameters and serious information loss caused by forward convolution and maximum pooling of a convolutional autoencoder and other convolution-based denoising autoencoders. Meanwhile, the feature map is adaptively encoded, and those feature maps containing target information are selectively emphasized, and those feature maps containing interference information are as much as possible ignored.

Description

Technical Field

[0001] This invention belongs to the technical fields of computer vision and deep learning, and relates to a CNN denoising method based on range-Doppler information, particularly a radar clutter suppression method. Background Technology

[0002] With the increasing number of radar sensors and the lack of standardized automotive radar frequency bands, mutual interference is inevitable and must be addressed. However, due to the complex environment in which targets are located and the limited improvement in clutter suppression performance, the processed signal may still contain residual strong clutter energy. In this situation, it will still cause significant problems for subsequent target detection. Therefore, how to quickly and accurately detect targets at high range under such conditions and improve the performance of target detection in strong clutter environments is an engineering problem that needs to be solved.

[0003] With the rapid development of deep learning technology, denoising autoencoders are being applied to anti-jamming in automotive radar. Normal convolutional neural networks are used to reduce noise in range-Doppler maps, while positive convolutional neural networks are used to denoise both the range-contour map and the range-Doppler map. Currently, most deep learning-based methods use simple fully convolutional networks as denoising autoencoders. However, considering the weight-sharing nature of convolution, it is meaningless to retain many features of noise and interference in the denoising result. Furthermore, in some cases, the signal-to-noise ratio of the range-Doppler map is too low to recover the target peak. Summary of the Invention

[0004] This invention provides a CNN denoising method based on distance-Doppler information to overcome the above-mentioned technical problems. The CNN denoising method based on distance-Doppler information includes the following steps: Acquire continuous multi-frame radar range-Doppler images, which include the current frame image, the previous frame image, and the two frames before the current frame image; A convolutional neural network model was constructed to denoise radar range-Doppler images; An effective loss function is introduced into the convolutional neural network model for training, resulting in a trained convolutional neural network model. The current frame image, the previous frame image, and the two frames before the current frame are input into the trained convolutional neural network model to obtain the denoised image of the current frame image.

[0005] Furthermore, the convolutional neural network model includes: Feature extraction module: used to extract features from the current frame image, the previous frame image, and the two frames before the current frame, and then fuse the extracted features; Encoder: Used to adaptively encode the fused features output by the feature extraction module, selectively emphasizing feature maps that contain target information and ignoring feature maps that contain interfering information. Decoder: Based on the encoded feature map output by the encoder, the feature map is adaptively decoded to obtain the denoised radar image.

[0006] The encoder and the decoder are connected using two hop connections.

[0007] Furthermore, the encoder includes: First encoder submodule: Based on the 9-channel 128×128 feature layer output by the feature extraction module, a 32-channel 64×64 feature layer is obtained; Second encoder submodule: Based on the 32-channel 64×64 feature layer output by the first encoder submodule, a 64-channel 32×32 feature layer is obtained; The third encoder submodule: Based on the 64-channel 32×32 feature layer output by the second encoder submodule, a 128-channel 16×16 feature layer is obtained.

[0008] Furthermore, the first encoder submodule, the second encoder submodule, and the third encoder submodule have the same structure; The first encoder submodule includes a stacked adaptive mobile encoder module and a first convolutional attention module.

[0009] Furthermore, the adaptive motion encoder module is divided into two parallel branches: a first parallel branch and a second parallel branch; The first parallel branch contains three parallel branches: the first branch, the second branch, and the third branch; The first branch consists of a 1 It consists of a convolution with a kernel size and a batch normalization operation connected sequentially; The second branch consists of a 3 It consists of sequentially connected asymmetric convolutional blocks of 3 kernel size and batch normalization operations; The third branch consists of a batch normalization operation; After this, the outputs of these three branches are summed element by element, and then the feature map is activated by the ReLU function; The second parallel branch includes two parallel branches: the fourth branch and the fifth branch; The fourth and fifth branches have the same structure; The fourth branch includes: a size of 3 The system consists of 3 asymmetric convolutional blocks with a stride of 2 and batch normalization operations connected sequentially to expand the feature dimension and receptive field. The output is then added element by element and activated by the ReLU function. Finally, the feature map is added to the result of the operation on the long residual edge from the original feature map, which contains a kernel size of 3. 3. Asymmetric convolution operation with a stride of 2.

[0010] Furthermore, the first convolutional block attention module is used to perform global max pooling and global average pooling on the input feature map, compressing the spatial information before inputting it into a shared multilayer perceptron to generate channel weight coefficients. These coefficients are used to select channel features that are more critical to the denoising task. Max pooling and average pooling are then performed along the channel axes on the channel-weighted feature map, and the resulting concatenated maps are processed by a 7-channel perceptron. The 7th convolutional layer generates a spatial weight map, which is used to focus the specific spatial location of the target in the radar spectrum and suppress background clutter areas.

[0011] Furthermore, the encoder also includes a channel expansion unit for expanding each frame of the input single-channel radar range-Doppler spectrum through a 1... The first convolutional layer performs linear projection, mapping the original single-channel data into a 3-channel feature vector while maintaining the feature map spatial resolution. This achieves a preliminary mapping from the low-dimensional signal space to the high-dimensional feature space. The three expanded images are then subjected to feature fusion, and the features from the three time steps are stitched together along the channel dimension.

[0012] Furthermore, the loss function is the mean squared error loss.

[0013] Furthermore, the decoder includes a first decoder submodule, a second decoder submodule, and a third decoder submodule connected in sequence; The first decoder submodule, the second decoder submodule, and the third decoder submodule have the same structure; The first decoder submodule includes a stacked adaptive moving decoder module and a second convolutional attention module.

[0014] A CNN denoising device based on distance-Doppler information includes: Acquisition module: used to acquire multiple consecutive frames of radar range-Doppler images, including the current frame image, the previous frame image, and the two frames before the current frame; Modules: Used to build convolutional neural network models for denoising radar range-Doppler images; Training module: used to train the convolutional neural network model by introducing an effective loss function, and to obtain a trained convolutional neural network model; The "Get" module is used to input the current frame image, the previous frame image, and the two frames before the current frame into the trained convolutional neural network model to obtain the denoised image of the current frame image.

[0015] Compared with existing technologies, the present invention has the following advantages and positive effects: This invention presents a CNN denoising method based on range-Doppler information. It employs an adaptive motion encoder and an adaptive motion decoder module, capturing temporal dependencies by inputting three consecutive frames of radar range-Doppler spectrum. An adaptive motion module is embedded in both the encoder and decoder, combining asymmetric convolutional blocks with depthwise separable convolutions to efficiently extract multi-scale spatial features while reducing the number of model parameters. A convolutional block attention module is introduced, using channel attention to calibrate feature channel weights and spatial attention to focus on key target regions, achieving adaptive suppression of clutter interference. This invention designs an adaptive mobile encoder module, combined with a convolutional block attention module, to effectively avoid the problems of redundant parameters and severe information loss caused by convolutional autoencoders and other convolutional denoising-based autoencoders that use forward convolution and max pooling. Simultaneously, it adaptively encodes feature maps, selectively emphasizing those containing target information and ignoring those containing interfering information as much as possible. An adaptive mobile decoder module is also designed, which has the same architecture as the mobile encoder module, the only difference being that all convolutional operations are replaced by deconvolution operations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be further introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is an overall flowchart of a CNN denoising method based on distance-Doppler information according to the present invention; Figure 2 This is a schematic diagram of the network structure of a CNN denoising method based on distance-Doppler information according to the present invention; Figure 3 This is a schematic diagram of the network structure of the adaptive mobile encoder of the present invention; Figure 4 This is a schematic diagram of the network structure of the adaptive mobile decoder of the present invention; Figure 5The diagram illustrates the radar range-Doppler spectrum denoising results of the simulation dataset applied to this invention; where (a) is the radar range-Doppler spectrum of the current frame; (b) is the radar range-Doppler spectrum of the previous frame; (c) is the radar range-Doppler spectrum of the previous two frames; (d) is the noise-free true value; and (e) is the prediction result after neural network processing. Figure 6 The diagram illustrates the denoising results of measured radar range-Doppler spectrum using this invention, where (a) is the radar range-Doppler spectrum of the current frame; (b) is the radar range-Doppler spectrum of the previous frame; (c) is the radar range-Doppler spectrum of the previous two frames; and (d) is the prediction result after processing by a neural network. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Figure 1 This is an overall flowchart of a CNN denoising method based on distance-Doppler information according to the present invention; This embodiment provides a CNN denoising method based on distance-Doppler information; S1: Acquire consecutive multi-frame radar range-Doppler images, which include the current frame image, the previous frame image, and the two frames before the current frame; S2: Construct a convolutional neural network model for denoising radar range-Doppler images; S3: Introduce an effective loss function to train the convolutional neural network model to obtain a trained convolutional neural network model; S4: Input the current frame image, the previous frame image, and the two frames before the current frame into the trained convolutional neural network model to obtain the denoised image of the current frame image.

[0020] The steps S1 / S2 / S3 / S4 are executed sequentially; A convolutional neural network (CNN) model is constructed for image feature extraction. The input is a series of consecutive radar range-Doppler images, and the output is a denoised radar range-Doppler image. The CNN model uses an adaptive mobile encoder and an adaptive mobile decoder module to adaptively encode and decode the feature maps. A convolutional block attention module selectively emphasizes feature maps containing target information and ignores feature maps containing interference information as much as possible. When the shape of the input feature layer is [C, H, W], the shape of the output feature layer is [2C, H / 2, W / 2]. Figure 2 This is a schematic diagram of the network structure of a CNN denoising method based on distance-Doppler information according to the present invention; The convolutional neural network model includes: Feature extraction module: used to extract features from the current frame image, the previous frame image, and the two frames before the current frame, and then fuse the extracted features; Encoder: Used to adaptively encode the fused features output by the feature extraction module, selectively emphasizing feature maps that contain target information and ignoring feature maps that contain interfering information. Decoder: Based on the encoded feature map output by the encoder, the feature map is adaptively decoded to obtain the denoised radar image.

[0021] The encoder and the decoder are connected using two hop connections.

[0022] The encoder includes: First encoder submodule: Based on the 9-channel 128×128 feature layer output by the feature extraction module, a 32-channel 64×64 feature layer is obtained; Second encoder submodule: Based on the 32-channel 64×64 feature layer output by the first encoder submodule, a 64-channel 32×32 feature layer is obtained; The third encoder submodule: Based on the 64-channel, 64×64 feature layer output by the second encoder submodule, a 128-channel, 16×16 feature layer is obtained.

[0023] The first encoder submodule, the second encoder submodule, and the third encoder submodule have the same structure; The first encoder submodule includes a stacked adaptive mobile encoder module and a first convolutional attention module.

[0024] Figure 3 This is a schematic diagram of the network structure of the adaptive mobile encoder of the present invention; The adaptive motion encoder module is divided into two parallel branches: the first parallel branch and the second parallel branch; The first parallel branch contains three parallel branches: the first branch, the second branch, and the third branch; The first branch consists of a 1 It consists of a convolution with a kernel size and a batch normalization operation connected sequentially; The second branch consists of a 3 It consists of sequentially connected asymmetric convolutional blocks of 3 kernel size and batch normalization operations; The third branch consists of a batch normalization operation; After this, the outputs of these three branches are summed element by element, and then the feature map is activated by the ReLU function; The second parallel branch includes two parallel branches: the fourth branch and the fifth branch; The fourth and fifth branches have the same structure; The fourth branch includes: a size of 3 The system consists of 3 asymmetric convolutional blocks with a stride of 2 and batch normalization operations connected sequentially to expand the feature dimension and receptive field. The output is then added element by element and activated by the ReLU function. Finally, the feature map is added to the result of the operation on the long residual edge from the original feature map, which contains a kernel size of 3. 3. Asymmetric convolution operation with a stride of 2.

[0025] The first convolutional block attention module performs global max pooling and global average pooling on the input feature map, compressing the spatial information before inputting it into a shared multilayer perceptron to generate channel weight coefficients. These coefficients are used to select channel features that are more critical to the denoising task. Max pooling and average pooling are then performed along the channel axes on the channel-weighted feature map, and the resulting concatenated maps are processed by a 7-channel array. The 7 convolutional layers generate spatial weight maps, which are used to focus on the specific spatial location of the target in the radar spectrum, suppress background clutter regions, selectively emphasize feature maps that contain target information, and ignore feature maps that contain interference information as much as possible.

[0026] Specifically, such as Figure 2 As shown, for the input of consecutive multi-frame radar range-Doppler images (current frame, previous frame, and two previous frames), the size is 128×128 with 1 channel. First, the images are stitched together in the channel dimension to form a multi-channel input tensor.

[0027] The input tensor is then encoded using an adaptive mobile encoder to avoid redundant parameters and severe information loss caused by forward convolution and max pooling in convolutional autoencoders and other convolution-based denoising autoencoders. Simultaneously, a convolutional attention module is employed to selectively emphasize feature maps containing target information while ignoring feature maps containing interfering information as much as possible.

[0028] The encoder also includes a channel expansion unit, which, for each frame of the input single-channel radar range-Doppler spectrum, expands the channel through a 1... The first convolutional layer performs linear projection, mapping the original single-channel data into a 3-channel feature vector while maintaining the feature map spatial resolution. This achieves a preliminary mapping from the low-dimensional signal space to the high-dimensional feature space. The three expanded images are then subjected to feature fusion, and the features from the three time steps are stitched together along the channel dimension.

[0029] The decoder includes a first decoder submodule, a second decoder submodule, and a third decoder submodule connected in sequence. The first decoder submodule, the second decoder submodule, and the third decoder submodule have the same structure; The first decoder submodule includes a stacked first adaptive moving decoder module and a second convolutional attention module; The adaptive motion encoder module and the adaptive motion decoder module have the same architecture; All convolution operations in the adaptive motion encoder module are replaced with deconvolution operations in the adaptive motion decoder module.

[0030] Figure 4 This is a schematic diagram of the network structure of the adaptive mobile decoder of the present invention; An adaptive moving decoder module is employed to adaptively decode the feature maps. Two hop connections are used between the encoder and decoder to mitigate gradient vanishing as the model trains. The adaptive moving decoder module has the same architecture as the moving encoder module. The only difference is that all convolutional operations are replaced by deconvolutional operations. When the input feature layer has a shape of [2C, H / 2, W / 2], the output feature layer has a shape of [C, H, W]. Specifically, the encoder outputs a feature layer with 128 channels and a size of 16×16. The first decoder submodule then obtains a 64-channel 32×32 feature layer, which is then summed with the 64-channel 32×32 feature layer output by the encoder. Finally, the second decoder submodule obtains a 32-channel 64×64 feature layer. Furthermore, the 32-channel 64×64 feature layer of the second decoder submodule is combined with the 32-channel 64×64 feature layer obtained from the encoder through a skip connection summation operation, and then the feature layer of 1 channel 128×128 is obtained through the third decoder module.

[0031] An effective loss function is introduced based on the radar range-Doppler spectrum denoising network model, and the network model is trained using the training dataset. Specifically, the loss function introduced in the radar range-Doppler spectrum denoising network model is the mean square error loss function. For each training sample, the mean of the sum of squared pixel-level Euclidean distances between the predicted spectrum output by the network and the clutter-free true spectrum is calculated. This loss value is used to guide the backpropagation of the network parameters' gradients, enabling the model to adaptively suppress background clutter and random noise while maintaining the target signal strength, ultimately achieving high signal-to-interference-plus-noise ratio target reconstruction.

[0032] A CNN denoising device based on distance-Doppler information includes: Acquisition module: used to acquire multiple consecutive frames of radar range-Doppler images, including the current frame image, the previous frame image, and the two frames before the current frame; Modules: Used to build convolutional neural network models for denoising radar range-Doppler images; Training module: used to train the convolutional neural network model by introducing an effective loss function, and to obtain a trained convolutional neural network model; The "Get" module is used to input the current frame image, the previous frame image, and the two frames before the current frame into the trained convolutional neural network model to obtain the denoised image of the current frame image.

[0033] Figure 5 As shown, the trained model can effectively denoise the radar range-Doppler spectrum and clearly acquire target signals.

[0034] Figure 5 The diagram illustrates the radar range-Doppler spectrum denoising results of the simulation dataset applied to this invention; where (a) is the radar range-Doppler spectrum of the current frame; (b) is the radar range-Doppler spectrum of the previous frame; (c) is the radar range-Doppler spectrum of the previous two frames; (d) is the noise-free true value; and (e) is the prediction result after neural network processing. Figure 6The diagram illustrates the denoising results of measured radar range-Doppler spectrum using this invention, where (a) is the radar range-Doppler spectrum of the current frame; (b) is the radar range-Doppler spectrum of the previous frame; (c) is the radar range-Doppler spectrum of the previous two frames; and (d) is the prediction result after processing by a neural network.

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A CNN denoising method based on distance-Doppler information, characterized in that, Includes the following steps: Acquire continuous multi-frame radar range-Doppler images, which include the current frame image, the previous frame image, and the two frames before the current frame image; A convolutional neural network model was constructed to denoise radar range-Doppler images; An effective loss function is introduced into the convolutional neural network model for training, resulting in a trained convolutional neural network model. The current frame image, the previous frame image, and the two frames before the current frame are input into the trained convolutional neural network model to obtain the denoised image of the current frame image.

2. The CNN denoising method based on distance-Doppler information according to claim 1, characterized in that, The convolutional neural network model includes: Feature extraction module: used to extract features from the current frame image, the previous frame image, and the two frames before the current frame, and then fuse the extracted features; Encoder: Used to adaptively encode the fused features output by the feature extraction module, selectively emphasizing feature maps that contain target information and ignoring feature maps that contain interfering information. Decoder: Based on the encoded feature map output by the encoder, the feature map is adaptively decoded to obtain the denoised radar image. The encoder and the decoder are connected using two hop connections.

3. The CNN denoising method based on distance-Doppler information according to claim 1, characterized in that the encoder... include: First encoder submodule: Based on the 9-channel 128×128 feature layer output by the feature extraction module, a 32-channel 64×64 feature layer is obtained; Second encoder submodule: Based on the 32-channel 64×64 feature layer output by the first encoder submodule, a 64-channel 32×32 feature layer is obtained; The third encoder submodule: Based on the 64-channel 32×32 feature layer output by the second encoder submodule, a 128-channel 16×16 feature layer is obtained.

4. The CNN denoising method based on distance-Doppler information according to claim 3, characterized in that, The first encoder submodule, the second encoder submodule, and the third encoder submodule have the same structure; The first encoder submodule includes a stacked adaptive mobile encoder module and a first convolutional attention module.

5. The CNN denoising method based on distance-Doppler information according to claim 4, characterized in that: The adaptive motion encoder module is divided into two parallel branches: the first parallel branch and the second parallel branch; The first parallel branch contains three parallel branches: the first branch, the second branch, and the third branch; The first branch consists of a 1 It consists of a convolution with a kernel size and a batch normalization operation connected sequentially; The second branch consists of a 3 It consists of sequentially connected asymmetric convolutional blocks of 3 kernel size and batch normalization operations; The third branch consists of a batch normalization operation; After this, the outputs of these three branches are summed element by element, and then the feature map is activated by the ReLU function; The second parallel branch includes two parallel branches: the fourth branch and the fifth branch; The fourth and fifth branches have the same structure; The fourth branch includes: a size of 3 The system consists of 3 asymmetric convolutional blocks with a stride of 2 and batch normalization operations connected sequentially to expand the feature dimension and receptive field. The output is then added element by element and activated by the ReLU function. Finally, the feature map is added to the result of the operation on the long residual edge from the original feature map, which contains a kernel size of 3.

3. Asymmetric convolution operation with a stride of 2.

6. The CNN denoising method based on distance-Doppler information according to claim 4, characterized in that: The first convolutional block attention module performs global max pooling and global average pooling on the input feature map, compressing the spatial information before inputting it into a shared multilayer perceptron to generate channel weight coefficients. These coefficients are used to select channel features that are more critical to the denoising task. Max pooling and average pooling are then performed along the channel axes on the channel-weighted feature map, and the resulting concatenated maps are processed by a 7-channel array. The 7th convolutional layer generates a spatial weight map, which is used to focus the specific spatial location of the target in the radar spectrum and suppress background clutter areas.

7. The CNN denoising method based on distance-Doppler information according to claim 4, characterized in that: The encoder also includes a channel expansion unit for expanding each input single-channel radar range-Doppler spectrum through a 1... The first convolutional layer performs linear projection, mapping the original single-channel data into a 3-channel feature vector while maintaining the feature map spatial resolution. This achieves a preliminary mapping from the low-dimensional signal space to the high-dimensional feature space. The three expanded images are then subjected to feature fusion, and the features from the three time steps are stitched together along the channel dimension.

8. The CNN denoising method based on distance-Doppler information according to claim 2, characterized in that: The loss function is the mean squared error loss.

9. A CNN denoising method based on distance-Doppler information according to claim 2, characterized in that: The decoder includes a first decoder submodule, a second decoder submodule, and a third decoder submodule connected in sequence. The first decoder submodule, the second decoder submodule, and the third decoder submodule have the same structure; The first decoder submodule includes a stacked adaptive moving decoder module and a second convolutional attention module.

10. A CNN denoising device based on distance-Doppler information, characterized in that, include: Acquisition module: used to acquire multiple consecutive frames of radar range-Doppler images, including the current frame image, the previous frame image, and the two frames before the current frame; Modules: Used to build convolutional neural network models for denoising radar range-Doppler images; Training module: used to train the convolutional neural network model by introducing an effective loss function, and to obtain a trained convolutional neural network model; The "Get" module is used to input the current frame image, the previous frame image, and the two frames before the current frame into the trained convolutional neural network model to obtain the denoised image of the current frame image.

Images