Radar image sea clutter suppression assisted target detection method and system

Abstract

The present invention discloses a radar image sea clutter suppression-assisted target detection method and system. The method comprises the following steps: 1) an image fusion module generates multiple image pairs, and each image pair is used as a training sample; the image pairs include a radar target image with a clean background fused to the radar target, and a radar image with a sea clutter background fused to the radar target; 2) the sea clutter radar image in the training sample is used as input, and the radar target image as supervision, and a conditional generative adversarial network is trained using the training sample; the conditional generative adversarial network includes a generator and a discriminator; 3) a radar image to be detected containing sea clutter is input into the generator of the trained conditional generative adversarial network to obtain a sea clutter-free image, which is then input into a radar target detection module based on deep learning to obtain a target detection result. The present invention improves the accuracy and reliability of target detection, while minimizing computational effort.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar image processing, and in particular relates to a radar image sea clutter suppression assisted target detection method and system based on a conditional generative adversarial network denoising algorithm. Background Art

[0002] Radar image processing technology processes radar signals to produce clear, visible radar images. Because radar systems often operate in extreme environments, such as inclement weather and electromagnetic interference, various interferences can appear in radar images, with sea clutter being a common one. This interference can negatively impact radar image visibility and target detection accuracy.

[0003] Traditional sea clutter suppression methods typically use filtering techniques or wavelet transforms, but these methods are not ideal and have certain limitations. With the development of deep learning technology, especially the emergence of generative adversarial networks (GANs) and conditional generative adversarial networks (CGANs), these technologies have been widely used in fields such as image denoising and image enhancement, and have achieved excellent results.

[0004] Traditional CGAN denoising algorithms primarily target visible light images in natural scenes. This paper attempts to employ the CGAN denoising algorithm to suppress sea clutter interference in radar images. This is particularly challenging because sea clutter is significantly affected by the environment, leading to diverse imaging results. We also introduced an innovative data augmentation algorithm to simulate its imaging results, enabling CGAN to perform well in suppressing sea clutter. This algorithm is then combined with corresponding detection algorithms to improve the accuracy of radar target detection. Summary of the Invention

[0005] In view of the limitations of the existing technology, the object of the present invention is to provide a radar image sea clutter suppression assisted target detection method and system using a conditional generative adversarial network (CGAN) denoising algorithm.

[0006] The technical solution adopted by the present invention is:

[0007] A radar image sea clutter suppression assisted target detection method, the steps of which include:

[0008] 1) The image fusion module generates multiple image pairs and uses each image pair as a training sample; the image pair includes a radar target image with a clean background fused with the radar target and a sea clutter radar image with a sea clutter background fused with the radar target;

[0009] 2) using the sea clutter radar image in the training sample as input and the radar target image as supervision, and using the training sample to train a conditional generative adversarial network; the conditional generative adversarial network includes a generator and a discriminator;

[0010] 3) A radar image to be detected containing sea clutter is input into the generator of the trained conditional generative adversarial network to obtain an image without sea clutter and input it into the deep learning-based radar target detection module to obtain the target detection result.

[0011] Furthermore, the method for obtaining the target detection result is as follows: first, the radar target detection module sequentially uses the convolution layer and the pooling layer to extract features from the input sea clutter-removed image; then the extracted features are sequentially passed through the residual layer and the feature pyramid network to extract feature maps of different scales; then the feature map of each scale is sequentially processed through the convolution layer and the fully connected layer to obtain the position and category information of the target in the feature map of the corresponding scale; and then the target detection result is presented in a visual manner.

[0012] Furthermore, through X ij =RandomMask(I ij )+γO ij Generate the radar target image; wherein X ij is the element in row i and column j in radar target image X, I ij is the element in row i and column j of the clean background image I, O ij is the element in the i-th row and j-th column of the radar target image O, γ is the weight coefficient, which is used to adjust the brightness of the radar target in the clean background image I; RandomMask is a mask image whose size is the size of the clean background image I, which is used to randomly select an area the size of the radar target image from the clean background image I for the fusion of the radar target area.

[0013] Furthermore, through X ij ’ =RandomMask(I ij ’ )+γO ij Generate the sea clutter radar image; wherein X ij ’ is the sea clutter radar image X ’ The element in row i and column j in ij Sea clutter background image I ’ The element in row i and column j in O ij is the element in row i and column j of radar target image O, and γ is the weight coefficient used to adjust the radar target in the sea clutter background image I ’The brightness in; RandomMask is a mask map whose size is the sea clutter background image I ’ The size of the background image I ’ A region of the size of the radar target image is randomly selected from the image for fusion of the radar target region.

[0014] Furthermore, in the two images of the same image pair, the radar target is the same but the background is different.

[0015] Furthermore, the generator is a U-shaped backbone network.

[0016] A radar image sea clutter suppression assisted target detection system, characterized by comprising an image fusion module, a denoising module generation module and a target detection module;

[0017] The image fusion module is configured to generate a plurality of image pairs, wherein the image pairs include a radar target image having a clean background fused with the radar target and a sea clutter radar image having a sea clutter background fused with the radar target;

[0018] The denoising module generation module is configured to use each image pair as a training sample, the sea clutter radar image in the training sample as input and the radar target image as supervision, and use the training sample to train a conditional generative adversarial network; the conditional generative adversarial network includes a generator and a discriminator; and use the generator in the trained conditional generative adversarial network as a denoising module;

[0019] The denoising module is used to denoise a radar image to be detected containing sea clutter to obtain an image free of sea clutter;

[0020] The target detection module is used to detect the obtained sea clutter-removed image to obtain a target detection result.

[0021] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the above method.

[0022] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0023] The radar image sea clutter suppression assisted target detection method based on the conditional generative adversarial network denoising algorithm of the present invention mainly includes three modules: an image fusion module of background and radar target, a denoising module using the conditional generative adversarial network, and a target detection module using deep learning;

[0024] The background and radar target image fusion module fuses the radar image with the background image and generates an image pair through a generator. The image pair includes a radar target image with a clean background and radar target and a sea clutter radar image with a sea clutter background and radar target. The radar image refers to an image with a radar target (such as a "ship"), and the background image refers to an image without a radar target and with noise and sea clutter.

[0025] A conditional generative adversarial network denoising module is used to process radar images containing sea clutter, remove sea clutter interference from the radar image, and generate a clear radar target image to improve the accuracy of subsequent detection and clear visualization.

[0026] The deep learning target detection module is used to perform target detection on the denoised radar target image and obtain the target detection result.

[0027] As mentioned above, a radar image sea clutter suppression assisted target detection method based on the conditional generative adversarial network denoising algorithm, in which in the generative conditional adversarial network denoising module, this method fuses the clean background and the target area image blocks in the radar image to generate a radar target image with a clean background, which is used as training data to train the conditional generative adversarial network, so that the subsequent sea clutter radar image can suppress sea clutter, thereby improving the recognition accuracy of the subsequent target detection module.

[0028] The advantages of the present invention are as follows:

[0029] The method of the present invention can more effectively suppress sea clutter interference in radar images and improve the accuracy and reliability of target detection. It also has the advantages of small computational complexity, high speed, and strong practicality, and is suitable for practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the algorithm flow chart of the radar image sea clutter suppression assisted target detection method based on the generative adversarial network denoising algorithm.

[0031] Figure 2 Figure 2 shows the fusion process of clean / sea clutter background and radar target images.

[0032] Figure 3 Flowchart for denoising based on conditional generative adversarial networks.

[0033] Figure 4 This is a flowchart of the radar target detection algorithm based on deep learning. DETAILED DESCRIPTION

[0034] The present invention will be described in further detail below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0035] like Figure 1 As shown in the figure, a radar image sea clutter suppression-assisted target detection method based on a generative adversarial network denoising algorithm consists of three main modules: a background and radar target image fusion module, a conditional generative adversarial network denoising module, and a deep learning-based radar target detection module. The background and radar target image fusion module acquires clean radar background images and radar background images with sea clutter, respectively, and fuses them with manually annotated or partially extracted targets from previous target detection to produce a large number of paired fused image training data. Paired data refers to two images: a radar target image consisting of a clean background and radar target, and a radar image with sea clutter consisting of a background image and radar target. These paired data can be used to train a conditional generative adversarial network using the output and generated results of a U-shaped denoising network. Paired data must be strongly correlated (i.e., within a pair of data, the radar target is the same but the background is different). A pair of data serves as a training sample, with a sea clutter radar image as input and a radar target image as output, to enable the network to fit the denoising process. The generative adversarial network consists of a discriminator and a generator. The generator is primarily based on a U-shaped backbone network. Conditional generative adversarial networks (GANs) are conditional networks that accept non-empty inputs (non-conditional GANs can be understood as random generation without any input). The input is an image with a sea clutter background. The goal is to fit the sea clutter image with a clean background, thereby completing the sea clutter suppression process. Once training is complete, the inference phase begins. The sea clutter radar image with a sea clutter background is fed into the trained U-shaped denoising network to produce a denoised image. The subsequent object detection framework directly applies this denoised radar image to output detected object boxes.

[0036] In order to train a conditional generative adversarial network so that its generator can suppress sea clutter in subsequent detection, this method collects some samples to create image pairs. Figure 2 This method demonstrates that it first collects radar images with clean backgrounds and radar images with significant sea clutter. Radar targets extracted through manual annotation or target detection algorithms are then fused to obtain a fused radar target image with a clean background. Similarly, this fusion method can be used to obtain radar images with sea clutter, resulting in a pair of data that serves as a training sample. The specific fusion formula is shown in Formula 1, where X is the fused image, I is the background image, O is the radar target image, i and j represent the i-th row and j-th column in the image, respectively, and γ is a weight coefficient used to adjust the target's brightness against the background. RandomMask is a mask image whose size is the same as the background image. It is primarily used to randomly select an area the size of the radar target image for fusion within the radar target region, increasing the diversity of radar target positions and preventing overfitting of the denoising model.

[0037] Xij =RandomMask(I ij )+γO ij (Formula 1)

[0038] The denoising module uses a conditional generative adversarial network (GAN). Paired data obtained from the background and target fusion modules (radar target images fused against a clean background and sea clutter radar images fused against a sea clutter background) are used as training data. A generator generates clean images, while a discriminator determines whether the generated images are realistic. After training, only the generator is used to generate clean radar target images from sea clutter noise images. These are then used for radar target detection in the subsequent deep learning radar target detection module. The generation (sea clutter suppression) aids understanding (radar target detection). During the inference phase, the discriminator is discarded, and only the generator is used to predict sea clutter. Specifically, a radar image with sea clutter is fed into the generator, which then generates an image without or with sea clutter removed, achieving sea clutter suppression.

[0039] An actual radar image containing sea clutter is input into the generator of the trained conditional generative adversarial network to obtain an image without sea clutter and input it into the radar target detection module based on deep learning.

[0040] The processing flow of the radar target detection module based on deep learning is as follows Figure 4 As shown in the figure, first, the radar image after removing sea clutter is fed into the neural network model as the input of the algorithm. The input image undergoes feature extraction through a series of convolutional layers and pooling layers. These layers help capture various features in the image, such as edges and textures. Next, residual connections and feature pyramid networks are used to extract feature maps of different scales. This allows for the detection of objects of different sizes at different levels. Convolutional and fully connected layers are applied to the feature maps to obtain the location and category information of the object. These layers transform the feature maps into outputs containing object boxes and category probabilities. The output is post-processed, including using non-maximum suppression (NMS) to remove redundant detection boxes, select the boxes with the highest confidence, and eliminate boxes with large overlaps. Finally, the output detection results are presented in a visual way. Boxes and category labels can be drawn on the image to show the objects detected by the algorithm.

[0041] The present invention adopts a lighter-weight network structure. By introducing lightweight modules and more efficient network design, it significantly reduces the computational complexity and model size while maintaining accuracy, enabling fast and real-time target detection even in resource-limited environments.

[0042] This paper optimizes single-stage detectors: Compared with traditional two-stage detectors (such as Faster R-CNN), the detection algorithm of this paper improves the accuracy and robustness of small object detection by introducing a more sophisticated feature pyramid network and a multi-scale training strategy.

[0043] The detection algorithm of the present invention adopts an adaptive model scaling method, which automatically adjusts the scale of the model according to the size of the target. It can better adapt to the detection needs of targets of different scales and maintain high accuracy when detecting small targets.

[0044] This paper adopts a multi-cascade detection strategy, dividing the network into different stages and performing object detection in each stage. This strategy can capture features of different scales and semantic levels at different levels, improving the precision and recall of object detection.

[0045] This method designs a specific backbone network to adapt to its downstream detection tasks. The network consists of a FOUS structure and a CSP structure. This method scales the radar pseudo-color image to a size of 640x640x3 as the network input. The input image is sliced through the FOUS layer, converting the resulting image into an image containing 12 channels. A convolution operation is then performed to obtain a double-sampled feature map with no information loss, thereby improving the computational speed of subsequent operations. Next, this method inputs the image that has passed through the FOUS layer into the CSP layer, which consists of a convolution operation and a residual unit. This is followed by a CSP layer, which consists of a convolution layer, a batch normalization (BN) layer, and an activation layer (SiLU). Multiple CSP and CBS layers generate feature maps of different sizes. These feature maps of different sizes are fed into the Neck module, where they are concatenated with related features in the channel dimension. They are then passed through a convolution module to obtain feature maps of different scales, which are used to predict the output detection results. During the pre-training phase, this method replaces the prediction module for detection results with a fully connected layer. Feature maps of different scales output by the Neck module are upsampled and concatenated in the channel dimension, followed by a pooling operation. The entire backbone network is optimized through gradient backpropagation. The subsequent detection head calculates the detection bounding box and ultimately outputs the detection bounding box result.

[0046] Some parts of the present invention are well known to those skilled in the art and are not described in detail.

[0047] While specific embodiments of the present invention have been disclosed for illustrative purposes, intended to facilitate understanding and implementation of the present invention, those skilled in the art will appreciate that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments, and the scope of protection claimed in the present invention shall be determined by the scope of the claims.

Claims

1. A radar image sea clutter suppression assisted target detection method, the steps comprising: 1) The image fusion module generates multiple image pairs and uses each image pair as a training sample; the image pair includes a radar target image with a clean background fused with the radar target and a sea clutter radar image with a sea clutter background fused with the radar target; in the two images of the same image pair, the radar target is the same but the background is different; wherein, by X ij =RandomMask(I ij )+γO ij Generate the radar target image; wherein X ij is the element in row i and column j in radar target image X, I ij is the element in row i and column j of the clean background image I, O ij is the element in the i-th row and j-th column of the radar target image O, γ is the weight coefficient used to adjust the brightness of the radar target in the clean background image I; RandomMask is a mask image whose size is the size of the clean background image I, which is used to randomly select an area the size of the radar target image from the clean background image I for fusion of the radar target area; the sea clutter radar image is generated by the mask image; 2) The sea clutter radar image in the training sample is used as input and the radar target image is used as supervision, and the training sample is used to train a conditional generative adversarial network; the conditional generative adversarial network includes a generator and a discriminator; the generator is based on a U-shaped backbone network, including a FOUS structure and a CSP structure; the input image is sliced through the FOUS layer, and the resulting image is converted into an image containing 12 channels, and then a convolution operation is performed to obtain a double-sampled feature map with no information loss, which is input to the CSP layer, which consists of a convolution operation and a residual unit, and then connected to a C The BS layer consists of a convolutional layer, a batch normalization layer, and an activation layer. Multiple CSP and CBS layers generate feature maps of different sizes, which are fed into the Neck module. These maps are concatenated with related features in the channel dimension and then passed through a convolution module to obtain feature maps of different scales for predicting the output detection results. In the pre-training phase, the prediction module is replaced with a fully connected layer, and the feature maps of different scales output by the Neck module are upsampled and concatenated in the channel dimension. After pooling, the entire backbone network is optimized through gradient backpropagation. 3) Inputting a radar image to be detected containing sea clutter into the generator of the trained conditional generative adversarial network to obtain an image without sea clutter and inputting the image into a radar target detection module based on deep learning to obtain a target detection result, including: first, the radar target detection module sequentially uses a convolution layer and a pooling layer to extract features from the input image without sea clutter; then sequentially passes the extracted features through a residual layer and a feature pyramid network to extract feature maps of different scales; then, the feature map of each scale is sequentially processed through a convolution layer and a fully connected layer to obtain the position and category information of the target in the feature map of the corresponding scale; and then presenting the target detection result in a visual manner.

2. A radar image sea clutter suppression assisted target detection system, characterized in that: It includes image fusion module, denoising module generation module and target detection module; The image fusion module is used to generate multiple image pairs, each of which includes a radar target image with a clean background fused with the radar target and a sea clutter radar image with a sea clutter background fused with the radar target; in the two images of the same image pair, the radar target is the same but the background is different; wherein, by X ij =RandomMask(I ij )+γO ij Generate the radar target image; wherein X ij is the element in row i and column j in radar target image X, I ij is the element in row i and column j of the clean background image I, O ij is the element in the i-th row and j-th column of the radar target image O, γ is the weight coefficient used to adjust the brightness of the radar target in the clean background image I; RandomMask is a mask image whose size is the size of the clean background image I, which is used to randomly select an area the size of the radar target image from the clean background image I for fusion of the radar target area; the sea clutter radar image is generated by the mask image; The denoising module generation module is used to take each image pair as a training sample, take the sea clutter radar image in the training sample as input and the radar target image as supervision, and use the training sample to train a conditional generative adversarial network; the conditional generative adversarial network includes a generator and a discriminator; the generator in the trained conditional generative adversarial network is used as a denoising module; the generator is based on a U-shaped backbone network, including a FOUS structure and a CSP structure; the input image is sliced through the FOUS layer, and the obtained image is converted into an image containing 12 channels, and then a convolution operation is performed to obtain a double-sampled feature map input C without information loss. The SP layer consists of a convolution operation and a residual unit, followed by a CBS layer, which consists of a convolution layer, a batch normalization layer, and an activation layer. Multiple CSP and CBS layers generate feature maps of different sizes, which are fed into the Neck module and concatenated with related features in the channel dimension. Feature maps of different scales are then passed through a convolution module to predict the output detection results. In the pre-training phase, the prediction module is replaced with a fully connected layer, and the feature maps of different scales output by the Neck module are upsampled and concatenated in the channel dimension. After a pooling operation, the entire backbone network is optimized through gradient backpropagation. The denoising module is used to denoise a radar image to be detected containing sea clutter to obtain an image free of sea clutter; The target detection module is used to detect the obtained sea clutter-removed image to obtain a target detection result, including: first, the target detection module sequentially uses a convolution layer and a pooling layer to extract features of the input sea clutter-removed image; then the extracted features are sequentially passed through a residual layer and a feature pyramid network to extract feature maps of different scales; then the feature map of each scale is sequentially processed through a convolution layer and a fully connected layer to obtain the position and category information of the target in the feature map of the corresponding scale; and then the target detection result is presented in a visual manner.

3. A server, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step of the method according to claim 1.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to claim 1 are implemented.

Images