SAR image target detection method based on multi-level Laplacian pyramid denoising

Abstract

The present invention relates to a SAR image target detection method based on multi-level Laplace pyramid denoising. The method comprises: decomposing the original image into different frequency domain subbands using a Laplace pyramid transform, suppressing noise in each high-frequency subband using a trainable threshold module, and obtaining a denoised image through Laplace pyramid reconstruction; concatenating the denoised image and the first Laplace transformed subband with the original image to fuse the spatial and frequency domain information of the image, forming a three-channel image, which is then input into a subsequent target detection network; and introducing an attention mechanism into the feature fusion module of the base network to assign different weights to each pixel in the feature map, thereby highlighting effective features. The method is adaptive and does not increase the network complexity or training difficulty due to changes in the input image scale. Generating a ground truth map containing scale information does not involve neural network training, resulting in low complexity and ease of operation. The generated ground truth map indicates the position and scale information of each target.

Description

Technical Field

[0001] The present invention belongs to the technical fields of deep learning, image processing and radar target detection, and in particular relates to a SAR image target detection method based on multi-level Laplacian pyramid denoising. Background Art

[0002] SAR target detection is a key aspect of high-resolution radar image interpretation. It has significant applications in disaster prevention and relief, emergency rescue, urban management, and ocean monitoring. With the recent rapid development of artificial intelligence, deep learning-based image processing technology has become the mainstream approach for intelligent recognition and detection of image targets. Researchers have proposed a series of SAR image target detection algorithms based on convolutional neural networks and verified their effectiveness.

[0003] Speckle noise is a significant factor affecting the accuracy of intelligent target recognition and detection in SAR images. It reduces image contrast, causing false or missed target detections, and impacting image interpretation. Traditional image denoising primarily involves spatial and frequency domain denoising. Intelligent image denoising methods are primarily based on deep learning, using the feature extraction capabilities of neural networks to extract effective features from the image and identify the noise. Image denoising based on these two methods has achieved excellent results, but some challenges remain:

[0004] Existing denoising methods serve only as a preprocessing step for image processing, failing to achieve end-to-end image processing. Denoising is often separated from subsequent image processing steps and is not coupled with the detection process, compromising the integrity of the solution and, consequently, the overall processing performance. Traditional frequency-domain denoising methods often set noise suppression thresholds based on empirical experience, which is susceptible to human experience and cannot achieve adaptability to each image, network, and image processing task. Therefore, it is crucial to construct an end-to-end target detection network with an autonomous denoising architecture to address the performance issues of synthetic aperture radar target detection in the presence of speckle noise. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of decoupling of denoising and detection processes during SAR image target detection under the influence of speckle noise. A SAR image target detection method based on multi-level Laplacian pyramid denoising is proposed. The method is based on the multi-level Laplacian pyramid denoising end-to-end SAR image target detection model: LPDNet.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0007] The SAR image target detection method is based on a multi-level Laplacian pyramid denoising end-to-end SAR image target detection model: LPDNet, which includes a denoising network, a feature fusion module, and a target detection network.

[0008] The denoising network is a multi-level Laplacian pyramid denoising network based on threshold adaptation; and the threshold adaptive multi-level Laplacian pyramid denoising network includes a Laplacian pyramid decomposition module, a threshold adaptive determination module and a Laplacian pyramid reconstruction module;

[0009] The threshold-adaptive multi-level Laplacian pyramid denoising method is based on the prior characteristic that speckle noise in most SAR images is concentrated in the high-frequency sub-band. The high-frequency sub-band image of the image is extracted through Laplacian pyramid decomposition and placed into a branch network composed of convolution and fully connected layers for supervised learning. The threshold is adaptively learned through training, and the high-frequency sub-band image is hard-thresholded according to the threshold. The high-frequency sub-band image is then combined with the low-frequency sub-band image and reconstructed through the Laplacian pyramid to obtain the denoised image. In this way, the threshold of Laplacian denoising is supervised by the training set label, which effectively avoids the degradation of recognition performance caused by empirically set thresholds.

[0010] The feature fusion module fuses the original image, the first-layer high-frequency subband image of the Laplacian pyramid decomposition, and the denoised image by stacking to obtain a fused image;

[0011] The target detection network is a target detection network based on the attention mechanism CBAM Yolox-tiny; wherein CBAM stands for Convolutional Block Attention Module;

[0012] The Yolox-tiny network includes a backbone network, a feature fusion part, and a detection head;

[0013] The backbone network includes multiple convolutional layers and pooling layers;

[0014] The SAR image target detection method based on multi-level Laplacian pyramid denoising relies on an end-to-end SAR image target detection model based on multi-level Laplacian pyramid denoising: LPDNet, which includes a denoising network, a feature fusion module, and a target detection network. The target detection method includes:

[0015] S1: Select the target detection dataset and divide it into training set, validation set and test set;

[0016] S1 The target detection dataset contains sample images and corresponding labels;

[0017] S2: Build a multi-level Laplacian pyramid denoising end-to-end SAR image target detection model, including:

[0018] S21, constructing a threshold-adaptive multi-level Laplacian pyramid denoising network, including constructing a Laplacian pyramid decomposition module, a threshold-adaptive determination module, and a Laplacian pyramid reconstruction module;

[0019] The input of the Laplacian pyramid decomposition module is the sample image, and the output is the high-frequency subband image and the low-frequency subband image obtained by Laplacian pyramid decomposition; the input of the threshold adaptive determination module is the high-frequency subband image obtained by the Laplacian pyramid decomposition module, and the output is the high-frequency subband image after denoising; the input of the Laplacian pyramid reconstruction module is the high-frequency subband image and the low-frequency subband image after denoising, and the output is the denoised image;

[0020] The low-frequency subband image obtained by Laplacian pyramid decomposition can be used as the input of the next-level Laplacian pyramid decomposition module. Through k-level cascade, k-level high-frequency subband images and low-frequency subband images are obtained. Then, the Laplacian pyramid decomposition, threshold adaptive determination, and Laplacian pyramid reconstruction process are repeated to form a threshold-adaptive multi-level Laplacian pyramid denoising network.

[0021] The value range of k is 0≤k≤4 and when k is equal to 0, it corresponds to one decomposition and the k-level high-frequency sub-band image is obtained by k+1 decompositions;

[0022] S22, construct feature fusion module;

[0023] S23, build an object detection network based on attention mechanism;

[0024] S3: Train the multi-level Laplacian pyramid denoising end-to-end SAR image target detection model to obtain the trained LPDNet, including:

[0025] S31: Initialize the network parameters, initialize the convolution kernel of the convolution layer and the weight value of each layer according to the set initialization value, and set a reasonable learning rate;

[0026] S32: Use the sample images and corresponding labels in the training set divided in S1 to train LPDNet, supervise the network parameters through the loss function, and output the trained LPDNet;

[0027] S4: Target detection phase, specifically including the following steps:

[0028] The test set divided in S1 is used for detection, the sample images in the test set are input into the trained LPDNet, and the target detection results are output.

[0029] The target detection dataset S1 is obtained by downloading the corresponding existing dataset according to the target detection task;

[0030] The target detection data set is divided into a training set, a validation set, and a test set, specifically: the target detection data set is divided according to a certain ratio;

[0031] The certain proportion is: the training set, validation set and test set account for x%, y%, z% respectively, and x%+y%+z%=1.

[0032] The feature fusion module S22 fuses the original image, the first-layer high-frequency sub-band image 1 of the Laplacian pyramid decomposition, and the denoised image by stacking to obtain a fused feature map;

[0033] Among them, fusion through stacking can effectively make up for the loss of image detail information caused by the denoising network.

[0034] S23 describes the introduction of the CBAM module into the feature fusion part of the Yolox-tiny network; the CBAM module includes multiple global pooling layers, maximum pooling layers, and convolutional layers; the fused feature map obtained by the feature fusion module is input into the target detection network based on the attention mechanism, that is, the Laplacian pyramid denoising network and the target detection network based on the attention mechanism are connected through the feature fusion part to complete the LPDNet construction.

[0035] Beneficial effects

[0036] The present invention relates to a target detection method based on Laplacian pyramid denoising threshold self-adaptation, which has the following beneficial effects:

[0037] 1. This method combines threshold-adaptive multi-level Laplacian pyramid denoising with an object detection network to achieve adaptive learning of the Laplacian denoising threshold, adapting to each image. This overcomes the drawback of traditional frequency-domain denoising, where the denoising threshold is set based on experience, which can affect both denoising and object detection performance.

[0038] 2. This method stacks the denoised image, the high-frequency subband image of the first layer of the Laplacian pyramid, and the original image to compensate for the details lost during the denoising process while enhancing the high-frequency information of the image, thereby improving the object detection effect.

[0039] 3. Add an attention module to the target detection network to increase the representation ability, focus on important features, suppress unnecessary features, and further enhance the target detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the overall network structure diagram of the multi-level Laplacian pyramid denoising end-to-end SAR image target detection method of the present invention, that is, the method flow chart;

[0041] Figure 2This is a diagram of the multi-level Laplacian pyramid denoising network structure based on threshold adaptive in Example 1 of the multi-level Laplacian pyramid denoising end-to-end SAR image target detection method of the present invention; the multi-level Laplacian pyramid denoising end-to-end SAR image target detection model is based on the multi-level Laplacian pyramid denoising network with threshold adaptive

[0042] Figure 3 It is a structural diagram of the attention mechanism module in the present invention;

[0043] Figure 4 This is the structural diagram of the CBAM Yolox-tiny target detection network based on the attention mechanism. DETAILED DESCRIPTION

[0044] The SAR image detection method based on Laplacian pyramid denoising threshold adaptation according to the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Example 1

[0046] This embodiment illustrates the specific implementation process of the method of the present invention for target detection on the SSDD dataset.

[0047] The SAR image target detection method based on multi-level Laplacian pyramid denoising relies on an end-to-end SAR image target detection model based on multi-level Laplacian pyramid denoising, LPDNet, which includes a denoising network, a feature fusion module, and a target detection network. The target detection method includes:

[0048] S1: Select the target detection dataset and divide it into training set, validation set and test set;

[0049] S1 The target detection dataset contains sample images and corresponding labels;

[0050] S2: Build a multi-level Laplacian pyramid denoising end-to-end SAR image target detection model, including:

[0051] S21, constructing a threshold-adaptive multi-level Laplacian pyramid denoising network, including constructing a Laplacian pyramid decomposition module, a threshold-adaptive determination module, and a Laplacian pyramid reconstruction module;

[0052] The input of the Laplacian pyramid decomposition module is the sample image, and the output is the high-frequency subband image and the low-frequency subband image obtained by Laplacian pyramid decomposition; the input of the threshold adaptive determination module is the high-frequency subband image obtained by the Laplacian pyramid decomposition module, and the output is the high-frequency subband image after denoising; the input of the Laplacian pyramid reconstruction module is the high-frequency subband image and the low-frequency subband image after denoising, and the output is the denoised image;

[0053] The low-frequency subband image obtained by Laplacian pyramid decomposition can be used as the input of the next-level Laplacian pyramid decomposition module. Through k-level cascade, k-level high-frequency subband images and low-frequency subband images are obtained. Then, the Laplacian pyramid decomposition, threshold adaptive determination, and Laplacian pyramid reconstruction process are repeated to form a threshold-adaptive multi-level Laplacian pyramid denoising network.

[0054] The value range of k is 0≤k≤4 and when k is equal to 0, it corresponds to one decomposition and the k-level high-frequency sub-band image is obtained by k+1 decompositions;

[0055] S22, construct feature fusion module;

[0056] S23, build an object detection network based on attention mechanism;

[0057] S3: Train the multi-level Laplacian pyramid denoising end-to-end SAR image target detection model to obtain the trained LPDNet, including:

[0058] S31: Initialize the network parameters, initialize the convolution kernel of the convolution layer and the weight value of each layer according to the set initialization value, and set a reasonable learning rate;

[0059] S32: Use the sample images and corresponding labels in the training set divided in S1 to train LPDNet, supervise the network parameters through the loss function, and output the trained LPDNet;

[0060] S4: Target detection phase, specifically including the following steps:

[0061] The test set divided in S1 is used for detection, the sample images in the test set are input into the trained LPDNet, and the target detection results are output.

[0062] The target detection dataset S1 is obtained by downloading the corresponding existing dataset according to the target detection task;

[0063] The target detection data set is divided into a training set, a validation set, and a test set, specifically: the target detection data set is divided according to a certain ratio;

[0064] The certain proportion is: the training set, validation set and test set account for x%, y%, z% respectively, and x%+y%+z%=1.

[0065] The feature fusion module S22 fuses the original image, the first-layer high-frequency sub-band image 1 of the Laplacian pyramid decomposition, and the denoised image by stacking to obtain a fused feature map;

[0066] Among them, fusion through stacking can effectively make up for the loss of image detail information caused by the denoising network.

[0067] S23 describes the introduction of the CBAM module into the feature fusion part of the Yolox-tiny network; the CBAM module includes multiple global pooling layers, maximum pooling layers, and convolutional layers; the fused feature map obtained by the feature fusion module is input into the target detection network based on the attention mechanism, that is, the Laplacian pyramid denoising network and the target detection network based on the attention mechanism are connected through the feature fusion part to complete the LPDNet construction.

[0068] The experimental data and configuration are as follows:

[0069] (1) Using the SSDD standard dataset, which contains 1160 SAR ship images of different resolutions and polarization modes, including 2358 ships of different scales and materials, it is a widely used dataset in the field of SAR image target detection.

[0070] (2) The size of the images in the dataset is about 500×500. To facilitate the experiment, the image size is unified to 416×416;

[0071] (3) The dataset is divided into training set, test set, and validation set in a ratio of 7:2:1;

[0072] (4) The computing platform is a Linux server equipped with an NVIDIA K80 GPU, and the model is implemented using the PyTorch framework.

[0073] Figure 1 The overall network structure diagram of the multi-level Laplacian pyramid denoising end-to-end SAR image target detection method is shown, that is, the flow diagram of the method. Among them, the CBAM Yolox-tiny target detection network is a target detection network that integrates the CBAM module and Yolox-tiny, that is, Figure 4 .

[0074] Figure 2 The structure diagram of the multi-level Laplacian pyramid denoising network based on threshold adaptation is shown in this example;

[0075] Figure 3 Shows the structure diagram of the attention mechanism module;

[0076] Figure 4 The structure diagram of the CBAM Yolox-tiny object detection network based on the attention mechanism in this example is shown.

[0077] The following is the specific implementation process. LPDNet has two stages: network construction and training stage and target detection stage:

[0078] Among them, the network construction and training phase mainly includes the following steps:

[0079] Step A.1: Construct a training sample set, specifically:

[0080] The images in the SSDD dataset were resized to 416×416 and randomly divided into training, test, and validation sets in a ratio of 7:2:1.

[0081] Step A.2: Construct a SAR image target detection network based on Laplacian pyramid denoising threshold adaptation, specifically:

[0082] According to the threshold adaptive multi-level Laplacian pyramid denoising network proposed in S2, it is constructed and combined with Figure 1 Specifically:

[0083] The input of the Laplacian pyramid decomposition module is the original image, and the output is the high-frequency subband image and the low-frequency subband image obtained by Laplacian pyramid decomposition; the input of the threshold adaptive determination module is the high-frequency subband image obtained by the Laplacian pyramid decomposition module, and the output is the denoised high-frequency subband image; the input of the Laplacian pyramid reconstruction module is the denoised high-frequency subband image and the low-frequency subband image, and the output is the denoised image;

[0084] The low-frequency subband image obtained by Laplacian pyramid decomposition can be used as the input of the next-level Laplacian pyramid decomposition module. By nesting the Laplacian pyramid decomposition three times, the first, second, and third-level high-frequency subband images and low-frequency subband images can be obtained. The threshold adaptive determination process is repeated for each level of high-frequency subband images. The denoised high-frequency subband image and the low-frequency subband image / the output image of the previous Laplacian pyramid reconstruction are input into the Laplacian pyramid reconstruction module to obtain the denoised image.

[0085] (Or this version: the image in the training set is input into the Laplacian pyramid decomposition module, and the output is the high-frequency sub-band image 1 and the low-frequency sub-band image 1 obtained by the Laplacian pyramid decomposition; the input of the threshold adaptive determination module 1 is the high-frequency sub-band image 1 obtained by the Laplacian pyramid decomposition module, and the output is the high-frequency sub-band image 1 after denoising; the low-frequency sub-band image 1 is subjected to Laplacian pyramid decomposition to obtain the high-frequency sub-band image 2 and the low-frequency sub-band image 2; the high-frequency sub-band image 2 is input into the threshold adaptive determination module 2 to obtain the high-frequency sub-band image 2*; the low-frequency sub-band image 2 is subjected to Laplacian pyramid decomposition to obtain the high-frequency sub-band image 2 Laplacian pyramid decomposition is performed to obtain a high-frequency subband image 3 and a low-frequency subband image 3; the high-frequency subband image 3 is input into a threshold adaptive determination module 3 to obtain a high-frequency subband image 3*; the high-frequency subband image 3* and the low-frequency subband image 3 are input into a Laplacian pyramid reconstruction module to obtain a reconstructed image 3* of the third layer; the reconstructed image 3* and the high-frequency subband image 2* are input into the Laplacian pyramid reconstruction module to obtain a reconstructed image 2* of the second layer; the reconstructed image 2* and the high-frequency subband image 1* are input into the Laplacian pyramid reconstruction module to obtain a reconstructed image 1* of the first layer, i.e., the denoised image;)

[0086] Among them, the threshold adaptive determination module is composed of convolutional layers and fully connected layers, as shown in the following example: Figure 2 As shown;

[0087] When step A.2 is implemented specifically: the Laplace pyramid transform is based on the Gaussian pyramid decomposition. Multi-resolution feature extraction of the image is achieved through Gaussian filtering and downsampling, and thumbnails of the original image of various resolutions are obtained to form a Gaussian pyramid. The high-resolution part of the image is at the bottom of the pyramid, and the low-resolution part is at the top of the pyramid. The essence of the Laplace pyramid is the detail component of the Gaussian pyramid image at the same level, which is obtained by subtracting the Gaussian pyramid images of two adjacent levels, that is, upsampling the image of the higher level to make it the same size as the image of the lower level, and then performing a differential operation with the image of the lower level through filtering. The obtained image is the Laplace pyramid image of the corresponding level, which can be expressed as:

[0088]

[0089] Among them, LP l represents the decomposed layer l image, G l is the l-th layer Gaussian pyramid image, It represents the result of upsampling and filtering the l+1th layer of Gaussian pyramid image, and N represents the highest level of Laplace decomposition.

[0090] The reconstruction process of the Laplacian pyramid is the opposite of the decomposition process. It starts from the top layer, goes through downsampling and addition, and iteratively calculates layer by layer until the bottom layer image. The process can be expressed as:

[0091]

[0092] When the threshold adaptive determination module in step A.2 is specifically implemented: a convolution operation is performed on the input image to obtain the output image of the layer;

[0093] Among them, the threshold adaptive determination module includes convolutional layer, fully connected layer, activation function, hard threshold processing part and parameter setting;

[0094] In this network, the convolution kernel size of the convolution layer is the same as the size of the input high-frequency sub-band image, and has a global receptive field of the high-frequency sub-band image, which can be expressed as:

[0095] O i =MaxPool(f i *I H +b i )

[0096] Among them, O i Represents the i-th feature channel of the convolutional layer output, MaxPool() is the maximum pooling function, f i is the convolution kernel corresponding to the input feature map and the i-th feature channel, I H is the high-frequency sub-band image input to the convolutional layer, b i Represents the bias of the convolutional layer. The fully connected layer takes the output of the convolutional layer as input, that is, O i , output a tensor D,D∈R 1×1×1 :

[0097] D=ReLU(WO+b)

[0098] Where ReLU() is the relu activation function. After the above operation, the denoising threshold corresponding to a high-frequency subband can be obtained. The noise in the high-frequency sub-image is suppressed by the hard threshold function, which is as follows:

[0099]

[0100] Among them, I H (x, y) represents the high-frequency subband coefficient after threshold processing, I(x, y) represents the high-frequency subband coefficient after Laplace pyramid decomposition, and threshold is the threshold determined by the convolutional neural network.

[0101] Step A.3: Feature fusion, specifically:

[0102] The image obtained by multi-level Laplacian pyramid denoising, the original image, and the first-level high-frequency sub-band image of the Laplacian pyramid transform are fused to compensate for the image information features lost by denoising while enhancing the image contour information. The fused feature image is then fed into the subsequent detection network.

[0103] Step A.4: Build an attention-based object detection network:

[0104] Among them, the CBAM module is as follows Figure 3 As shown in the figure, it includes channel attention mechanism and spatial attention mechanism, which derives attention map along the channel and spatial dimensions in sequence, multiplies it with the image, and adaptively refines the features, thereby increasing the representation ability, that is, focusing on important features and suppressing unnecessary features.

[0105] The CBAM module is introduced into the feature fusion part of the Yolox-tiny network structure as follows Figure 4 As shown in the figure, the box in the figure is the attention mechanism module. CBAM is placed on four channels, assigning different weights to features. After the channels are connected, it provides effective information for the feature extraction process of the next layer and suppresses invalid information, thereby further improving the detection accuracy.

[0106] At this point, the feature fusion module connects the threshold-adaptive multi-level Laplacian pyramid denoising network and the object detection network based on the attention mechanism to form an end-to-end detection network, namely LPDNet;

[0107] Among them, the final result of target detection is Figure 4 The detection head output in contains the location information and category information of the detected target;

[0108] Step A.5: Train LPDNet, that is, use the training set constructed in step A.1 to train the threshold and target position of target detection in the multi-level Laplacian pyramid network, specifically:

[0109] Step A.5.1: Initialize the network parameters. Initialize the convolution kernel of the convolution layer and the weight value of each layer to a mean of 0 and a variance of Gaussian distribution;

[0110] Step A.5.2: Use StepLR as the learning rate adjustment mechanism, i.e., after each epoch, the learning rate is reduced by a factor of 0.05; use Adam (Adaptive momentum) as the optimizer algorithm;

[0111] The model first freezes the backbone network for 50 rounds of training, with the batch size (batch size during network training) set to 8 and the initial learning rate set to 0.001; then unfreezes and trains for 100 rounds, with the batch size set to 4 and the initial learning rate set to 0.0001;

[0112] Among them, the loss function is set as:

[0113]

[0114] in, and They are classification loss and regression loss respectively. The classification loss uses binary cross entropy loss, and the regression loss uses IoU loss function;

[0115] At this point, from step A.1 to step A.5.2, the network construction and training phase of this embodiment is completed;

[0116] In the detection phase, the model trained with the training set is used to detect the corresponding test samples to obtain the detection result graph. The detection effect is evaluated using precision (Precision), recall rate (Recall), average precision (AP), and F1 score (F1-score) as evaluation indicators.

[0117] These evaluation criteria are calculated based on four components: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). In this paper, TP and TN represent the number of correctly detected ships and the number of correct backgrounds, respectively. FP represents the number of false positives, and FN represents the number of undetected ships. To determine whether the detected object is correct, the intersection over union (IoU) is introduced. IoU is calculated as the ratio of overlap between the detection box and the ground truth box:

[0118]

[0119] Among them, S ∩ Represents the overlapping area between the predicted box and the real box, S ∪ Represents the overall area of the two. If the IoU is greater than the set threshold (set to 0.5 in this article), the detection box is determined to be correct.

[0120] Accuracy, recall, average precision, and F1 score are defined as follows:

[0121]

[0122]

[0123]

[0124]

[0125] Accuracy represents the proportion of all predicted boxes that are correct, while recall represents the proportion of correctly located and identified objects out of the total number of objects. AP and F1 scores measure the balance between accuracy and recall. Accuracy is calculated as the average precision of recall in the range of 0 to 1, while F1 is the harmonic mean of the two. Higher AP and F1 values indicate better detection performance.

[0126] Table 1 Experimental results

[0127]

[0128] Thus, the specific implementation of the SAR image target detection method based on the multi-level Laplacian pyramid denoising network in SAR ship detection described in the present invention is completed.

[0129] The above description is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the accompanying drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A SAR image target detection method based on multi-level Laplacian pyramid denoising, characterized in that: LPDNet is an end-to-end SAR image target detection model based on multi-level Laplacian pyramid denoising, which includes a denoising network, a feature fusion module, and a target detection network. The target detection method comprises: S1: Select the target detection dataset and divide it into training set, validation set and test set; S1 The target detection dataset contains sample images and corresponding labels; S2: Build a multi-level Laplacian pyramid denoising end-to-end SAR image target detection model, including: S21, constructing a threshold-adaptive multi-level Laplacian pyramid denoising network, including constructing a Laplacian pyramid decomposition module, a threshold-adaptive determination module, and a Laplacian pyramid reconstruction module; The input of the Laplacian pyramid decomposition module is the sample image, and the output is the high-frequency subband image and the low-frequency subband image obtained by Laplacian pyramid decomposition; the input of the threshold adaptive determination module is the high-frequency subband image obtained by the Laplacian pyramid decomposition module, and the output is the high-frequency subband image after denoising; the input of the Laplacian pyramid reconstruction module is the high-frequency subband image and the low-frequency subband image after denoising, and the output is the denoised image; The low-frequency subband image obtained by Laplacian pyramid decomposition can be used as the input of the next-level Laplacian pyramid decomposition module. Through k-level cascade, k-level high-frequency subband images and low-frequency subband images are obtained. Then, the Laplacian pyramid decomposition, threshold adaptive determination, and Laplacian pyramid reconstruction process are repeated to form a threshold-adaptive multi-level Laplacian pyramid denoising network. The value range of k is 0≤k≤4 and when k is equal to 0, it corresponds to one decomposition and the k-level high-frequency sub-band image is obtained by k+1 decompositions; S22, construct feature fusion module; S23, build an object detection network based on attention mechanism; S3: Train the multi-level Laplacian pyramid denoising end-to-end SAR image target detection model to obtain the trained LPDNet, including: S31: Initialize the network parameters, initialize the convolution kernel of the convolution layer and the weight value of each layer according to the set initialization value, and set a reasonable learning rate; S32: Use the sample images and corresponding labels in the training set divided in S1 to train LPDNet, supervise the network parameters through the loss function, and output the trained LPDNet; S4: Target detection phase, specifically including the following steps: The test set divided in S1 is used for detection, the sample images in the test set are input into the trained LPDNet, and the target detection results are output.

2. The SAR image target detection method based on multi-level Laplacian pyramid denoising according to claim 1, characterized in that: The target detection dataset S1 is obtained by downloading the corresponding existing dataset according to the target detection task; The target detection data set is divided into a training set, a validation set, and a test set, specifically: the target detection data set is divided according to a certain ratio; The certain proportion is: the training set, validation set and test set account for x%, y%, z% respectively, and x%+y%+z%=1.

3. The SAR image target detection method based on multi-level Laplacian pyramid denoising according to claim 2, characterized in that: The feature fusion module S22 fuses the original image, the first-layer high-frequency sub-band image 1 of the Laplacian pyramid decomposition, and the denoised image by stacking to obtain a fused feature map; Among them, fusion through stacking can effectively make up for the loss of image detail information caused by the denoising network.

4. The SAR image target detection method based on multi-level Laplacian pyramid denoising according to claim 3, characterized in that: S23 introduces the CBAM module into the feature fusion part of the Yolox-tiny network; the CBAM module includes multiple global pooling layers, maximum pooling layers, and convolutional layers; the fused feature map obtained by the feature fusion module is input into the target detection network based on the attention mechanism, that is, the Laplacian pyramid denoising network and the target detection network based on the attention mechanism are connected through the feature fusion part to complete the LPDNet.

Images