A SAR image target detection method and system based on point features

Abstract

The present invention relates to a SAR image target detection method and system based on point features. The method includes: S1: Input the SAR image into a feature extraction module to obtain feature maps of different scales; S2: Input the feature maps into a point feature detection network for convolutional processing to extract point features; S3: According to the minimum and maximum x and y values in the point features, convert them into pseudo-detection boxes; S4: After performing deformable convolutional operations on the processed feature maps and point features, and then through 1×1 convolution, output the corrected boxes of the pseudo-detection boxes and the target categories; Add the corrected boxes of the pseudo-detection boxes and the pseudo-detection boxes to obtain the final predicted target bounding box; S5: Construct a total loss function L to train the point feature detection network; Step S6: Input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result. The method provided by the present invention is specifically designed for the discrete characteristics of the target, and has a fast detection speed and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and specifically relates to a method and system for SAR image target detection based on point features. Background Art

[0002] Synthetic Aperture Radar (SAR) is a radar that can obtain images all day and all weather, and is widely used in military and civilian fields. With the development of the field of intelligent image processing, the function of convolutional neural network is becoming increasingly powerful, and the detection method based on convolutional neural network has powerful performance in the detection of various targets. With the development of the SAR image field, a large number of SAR images can be obtained, and it has great value to use convolutional neural network to interpret SAR images.

[0003] Since in SAR images, targets usually show the characteristics of high discretization degree and strong pose variability, and the mainstream target detectors are all designed for detecting optical images. Facing the characteristics of strong target scattering and strong variability in SAR images, most existing detectors cannot adapt to SAR target detection, so better detection performance cannot be achieved.

[0004] In addition, most existing algorithms blindly stack the number of network layers and use new technologies, resulting in a large amount of computation and parameters. In actual application scenarios, detection algorithms are usually deployed on edge devices and mobile devices with limited computing power. Therefore, algorithms with large amounts of computation and parameters cannot adapt to the actual application scenarios of SAR image target detection. Therefore, in SAR target detection, reducing the number of algorithm parameters and the amount of computation is very efficient.

[0005] Finally, existing SAR image target detection algorithms are mainly based on the design of anchor boxes, but the parameter setting of prior anchor boxes is extremely dependent on expert knowledge, and the algorithm based on anchor boxes has limitations in the detection of small targets in SAR. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a method and system for SAR image target detection based on point features.

[0007] The technical solution of the present invention is: a method for SAR image target detection based on point features, including:

[0008] Step S1: Input the SAR image into the feature extraction module to obtain feature maps of different scales;

[0009] Step S2: Input the feature map into the point feature detection network for convolutional processing to obtain the processed feature map, and output the point feature through a 1×1 convolution of the processed feature map;

[0010] Step S3: Select the minimum and maximum x and y values among the point features as the coordinates of the upper left corner and the lower right corner of the pseudo-detection box, and convert the point features into a pseudo-detection box;

[0011] Step S4: Perform deformable convolutional DCNv2 operation on the processed feature map and the point features, and then pass the operation result through a 1×1 convolution to output the corrected box of the pseudo-detection box and the target category; Add the corrected box of the pseudo-detection box and the pseudo-detection box to obtain the final predicted target bounding box;

[0012] Step S5: Based on the predicted class loss L 类别 , the predicted target bounding box loss L 包围框 and the pseudo-detection box loss L 伪检测框 Construct the total loss function L to train the point feature detection network, and update the network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges, save the network structure and parameters, and obtain the trained point feature detection network;

[0013] Step S6: Input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. The present invention discloses a method for SAR image target detection based on point features. In view of the strong scattering characteristics of SAR targets, a clever design of using a discrete point set to describe SAR targets improves the detection accuracy of SAR targets.

[0016] 2. The present invention solves the problem that the setting of anchor box parameters in the prior art highly depends on expert experience, adopts a non-anchor box algorithm, and has better detection performance for small targets in SAR. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of a method for SAR image target detection based on point features in an embodiment of the present invention;

[0018] Figure 2 is a schematic diagram of the structure of the point feature detection network in an embodiment of the present invention;

[0019] Figure 3 is a visualization diagram of the spatial position of the point features output by the point feature detection network and the predicted target bounding box in an embodiment of the present invention;

[0020] Figure 4 is a block diagram of the structure of a SAR image target detection system based on point features in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] The present invention provides a SAR image target detection method based on point features, which is specifically designed for the discrete characteristics of targets and has a relatively fast detection speed and accuracy.

[0022] In order to make the objectives, technical solutions and advantages of the present invention clearer, the following further elaborates on the present invention through specific embodiments and in conjunction with the accompanying drawings.

[0023] Embodiment 1

[0024] As Figure 1 shown, a SAR image target detection method based on point features provided by an embodiment of the present invention includes the following steps:

[0025] Step S1: Input the SAR image into the feature extraction module to obtain feature maps of different scales;

[0026] Step S2: Input the feature map into the point feature detection network for convolution processing to obtain the processed feature map, and output the point features through a 1×1 convolution of the processed feature map;

[0027] Step S3: Select the minimum and maximum x and y values in the point features as the coordinates of the upper left corner and the lower right corner of the pseudo-detection box, and convert the point features into pseudo-detection boxes;

[0028] Step S4: Perform deformable convolution DCNv2 operation on the processed feature map and the point features, and then output the corrected box of the pseudo-detection box and the target category through a 1×1 convolution; add the corrected box of the pseudo-detection box and the pseudo-detection box to obtain the final predicted target bounding box;

[0029] Step S5: Based on the predicted class loss L 类别 、predicted target bounding box loss L 包围框 and pseudo-detection box loss L 伪检测框 Construct the total loss function L to train the point feature detection network, and update the network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges, save the network structure and parameters, and obtain the trained point feature detection network;

[0030] Step S6: Input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result.

[0031] In one embodiment, the above step S1: Input the SAR image into the feature extraction module to obtain feature maps of different scales, specifically including:

[0032] The feature extraction module includes: ResNet50 and FPN; First, the SAR image is input into ResNet50 and undergoes five downsamplings. Then, the feature maps from the last three downsamplings are fed into FPN for feature fusion, and three feature maps of different sizes are output for predicting targets at different scales.

[0033] In the embodiment of the present invention, the feature extraction module is composed of ResNet50 and FPN. The SAR image undergoes five downsamplings by ResNet50, and then the feature maps from the last three downsamplings are fed into FPN for feature fusion, and three feature maps of different sizes are output. The number of channels of the three feature maps output by FPN is 256 dimensions.

[0034] In one embodiment, the above step S2: performing convolutional processing on the feature map input to the point feature detection network to obtain the processed feature map, and outputting point features through a 1×1 convolution on the processed feature map, specifically includes:

[0035] Step S21: Processing the feature map through three convolutional layers to obtain the processed feature map;

[0036] As Figure 2 shown, the three feature maps of different sizes output in step S1 are further processed through three 3×3 convolutional layers. During the three convolutional operations, the width, height, and number of channels of the feature map remain unchanged.

[0037] Step S22: Passing the processed feature map through a 1×1 convolutional layer to output a 3×N-dimensional point feature matrix, where N represents the number of discrete points in the point set; the first dimension is the coordinate displacement of the discrete point relative to the center point on the x-axis, the second dimension is the coordinate displacement of the discrete point relative to the center point on the y-axis, and the third dimension is the weight value of each of the N points, respectively representing the importance of each point.

[0038] Pass the processed feature map output in step S21 through a branch of a 1×1 convolution. After passing through the 1×1 convolution, the 256-dimensional feature map becomes a 3×N-dimensional point feature. The point feature adaptively learns the key semantic points of the target and describes the structure and position of the target in the feature map. In the 3×N-dimensional point feature matrix, a 3×N-dimensional point feature matrix is output, where N represents the number of discrete points in the point set; the first dimension is the coordinate displacement of the discrete point relative to the center point on the x-axis, the second dimension is the coordinate displacement of the discrete point relative to the center point on the y-axis, and the third dimension is the weight value of each of the N points, respectively representing the importance of each point.

[0039] Through iterative training, the N discrete points represented by the point features adaptively learn the position and structural information of the target. In the embodiments of the present invention, a set containing N discrete points is output for each target in the input SAR image, and the position and structure of the target are symbolized and described by the set of discrete points.

[0040] In one embodiment, in the above step S3: The minimum and maximum x and y values in the point features are selected as the coordinates of the upper left corner and the lower right corner of the pseudo-detection box, and the point features are converted into a pseudo-detection box.

[0041] According to the point features obtained in step S2, based on their coordinate values, a pseudo-detection box that can contain all N discrete points is determined.

[0042] In one embodiment, in the above step S4: The processed feature map and the point features are subjected to deformable convolution DCNv2 operation, and the operation result is then passed through a 1×1 convolution to output the corrected box of the pseudo-detection box and the target category; the corrected box of the pseudo-detection box and the pseudo-detection box are added to obtain the final predicted target bounding box, which specifically includes:

[0043] The point features and the processed features output in step S21 Figure 1 are subjected to deformable convolution DCNv2 operation. At this time, the point features represent the sampling position when the convolution kernel operates on the point of the feature map and the summation weight of each point. In this process, the point features adaptively model the scattering features of the target and provide a more efficient feature map for result prediction. In the deformable convolution DCNv2 operation, the width and height of the feature map remain unchanged, and the number of channels is 256 dimensions; then a ReLu activation operation is performed to activate the operation result of the deformable convolution DCNv2. Finally, the activation result is passed through a 1×1 convolution to output a 4+C-dimensional prediction vector; among them, the predicted corrected box of the pseudo-detection box is 4-dimensional, and the target category information is C-dimensional. C is the number of target categories, and the C dimensions respectively represent the confidence levels of different categories, and the category represented by the dimension with the largest value is used as the predicted category.

[0044] Finally, the corrected box of the pseudo-detection box in this step and the pseudo-detection box obtained in step S3 are added to obtain the final target bounding box.

[0045] In one embodiment, in the above step S5: Based on the predicted category loss L 类别 、predicted target bounding box loss L 包围框 and pseudo-detection box loss L 伪检测框 a total loss function L is constructed to train the point feature detection network, and the network parameters are updated through backpropagation. This step is repeated until the performance of the point feature detection network converges, the network structure and parameters are saved, and a trained point feature detection network is obtained, which specifically includes:

[0046] Construct the loss function: predicted category loss L类别 , the predicted target bounding box loss L 包围框 and the predicted false detection box loss L 伪检测框 , and the calculation formula for each part is as follows; predict represents the predicted output value, and GT represents the true label value; F 包围框 , F 伪检测框 and F 类别 represent loss calculation functions, which are SmoothL1, SmoothL1, and FocalLoss functions respectively:

[0047] L 包围框 = F 包围框 (predict 包围框 , GT 包围框 )

[0048] L 伪检测框 = F 伪检测框 (predict 伪检测框 , GT 包围框 )

[0049] L 类别 = F 类别 (predict 类别 , GT 类别 )

[0050] Construct the total loss function:

[0051] L = μ1 * L 类别 + μ2 * L 包围框 + μ3 * L 伪检测框

[0052] Among them, μ1, μ2, and μ3 are the preset weights of the three loss functions, which are 1.0, 1.0, and 0.5 respectively in the embodiments of the present invention.

[0053] Since the point features participate in the deformable convolution DCNv2 calculation, the point features are supervised by the indirect class loss, the target bounding box loss, and the direct false detection box loss, so that the discrete points in the point features can learn the information of the target key semantic points.

[0054] In one embodiment, the above step S6: Input the SAR image to be detected into the trained point feature detection network, and obtain the target bounding box and target category information as the detection result.

[0055] As Figure 3 shown, it shows the visual results of the spatial position of the point features (a set of 9 discrete points) obtained by the point feature detection network and the predicted target bounding box.

[0056] The present invention discloses a method for SAR image target detection based on point features. In view of the strong scattering characteristics of SAR targets, a clever design using a discrete point set to describe SAR targets is adopted, which improves the detection accuracy of SAR targets. The present invention solves the problem that the setting of anchor box parameters in the prior art highly depends on expert experience, and adopts a non-anchor box algorithm, which has better detection performance for small targets in SAR images.

[0057] Embodiment 2

[0058] As Figure 4 shown, the embodiment of the present invention provides a SAR image target detection system based on point features, including the following modules:

[0059] Feature extraction module 71, which is used to input the SAR image into the feature extraction module to obtain feature maps of different scales;

[0060] Point feature extraction module 72, which is used to perform convolutional processing on the feature map by inputting it into a point feature detection network, obtain the processed feature map, and output point features through a 1×1 convolution of the processed feature map;

[0061] Pseudo detection box generation module 73, which is used to select the minimum and maximum x and y values in the point features as the coordinates of the upper left corner and the lower right corner of the pseudo detection box, and convert the point features into pseudo detection boxes;

[0062] Predicted target bounding box module 74, which is used to perform deformable convolution DCNv2 operation on the processed feature map and the point features, and then output the corrected box of the pseudo detection box and the target category through a 1×1 convolution; add the corrected box of the pseudo detection box and the pseudo detection box to obtain the final predicted target bounding box;

[0063] Loss function construction module 75, which is used to construct the total loss function L based on the predicted class loss L 类别 , predicted target bounding box loss L 包围框 and pseudo detection box loss L 伪检测框 to train the point feature detection network, and update the network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges, save the network structure and parameters, and obtain the trained point feature detection network;

[0064] Detection module 76, which is used to input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result.

[0065] The above embodiments are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. All equivalent substitutions and modifications made without departing from the spirit and principle of the present invention shall be covered within the scope of the present invention.

Claims

1. A SAR image target detection method based on point features, characterized in that, Including: Step S1: Input the SAR image into the feature extraction module to obtain feature maps of different scales; Step S2: Input the feature map into the point feature detection network for convolution processing to obtain the processed feature map, and output the point feature through a 1×1 convolution of the processed feature map; Step S3: Select the minimum and maximum x and y values among the point features as the coordinates of the upper left and lower right corners of the pseudo-detection box, and convert the point features into pseudo-detection boxes; Step S4: Perform deformable convolution DCNv2 operation on the processed feature map and the point feature, and then perform a 1×1 convolution on the operation result to output the pseudo-detection box correction box and the target category; Add the pseudo-detection box correction box and the pseudo-detection box to obtain the final predicted target bounding box; Step S5: Based on the predicted class loss L 类别 , the predicted target bounding box loss L 包围框 and the pseudo-detection box loss L 伪检测框 Construct the total loss function L to train the point feature detection network, and update the network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges. Save the network structure and parameters to obtain the trained point feature detection network; Step S6: Input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result.

2. The SAR image target detection method based on point features according to claim 1, characterized in that The step S1: Input the SAR image into the feature extraction module to obtain feature maps of different scales, specifically including: The feature extraction module includes: ResNet50 and FPN; first, input the SAR image into ResNet50, perform five downsamplings, and then input the feature maps of the last three downsamplings into FPN for feature fusion, and output three feature maps of different sizes for predicting targets of different scales.

3. The SAR image target detection method based on point features according to claim 2, wherein, The step S2: Input the feature map into the point feature detection network for convolution processing to obtain the processed feature map, and output the point feature through a 1×1 convolution of the processed feature map, specifically including: Step S21: Process the feature map through three convolutional layers to obtain the processed feature map; Step S22: Output a 3×N-dimensional point feature matrix through a 1×1 convolutional layer, where N represents the number of discrete points in the point set; the first dimension is the coordinate displacement of the discrete point relative to the center point on the x-axis, the second dimension is the coordinate displacement of the discrete point relative to the center point on the y-axis, and the third dimension is the weight value of each of the N points, respectively representing the importance of each point.

4. The SAR image target detection method based on point features according to claim 3, wherein, The step S4: Perform deformable convolution DCNv2 operation on the processed feature map and the point feature, and then perform a 1×1 convolution on the operation result to output the pseudo-detection box correction box and the target category; Adding the pseudo-detection box correction box and the pseudo-detection box to obtain the final predicted target bounding box, specifically including: Perform scattering feature sampling on the processed feature map and the point feature by deformable convolution DCNv2, and then perform a 1×1 convolution to output a 4+C-dimensional prediction vector; among them, the predicted pseudo-detection box correction box is 4-dimensional, and the target category information is C-dimensional, C is the number of target categories, and the C dimensions respectively represent the confidence levels of each different category, and the category represented by the dimension with the largest value is used as the predicted category; in the deformable convolution DCNv2 operation, the point feature represents the sampling position when the convolution kernel reaches this point of the feature map and the summation weight of each sampling point.

5. The SAR image target detection method based on point features according to claim 4, wherein The step S5: Based on the predicted class loss L 类别 , the predicted target bounding box loss L 包围框 and the pseudo-detection box loss L 伪检测框 Construct the total loss function L to train the point feature detection network, and update the network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges. Save the network structure and parameters to obtain the trained point feature detection network. Specifically, it includes: Construct the loss function: the predicted class loss L 类别 , the predicted target bounding box loss L 包围框 and the predicted false detection box loss L 伪检测框 , the calculation formulas for each part are as follows; predict represents the predicted output value, and GT represents the true label value; F 包围框 , F 伪检测框 and F 类别 represent the loss calculation functions, which are the SmoothL1, SmoothL1, and FocalLoss functions respectively: L 包围框 = F 包围框 (predict 包围框 , GT 包围框 ) L 伪检测框 = F 伪检测框 (predict 伪检测框 , GT 包围框 ) L 类别 = F 类别 (predict 类别 , GT 类别 ) Construct the total loss function: L = μ1 * L 类别 + μ2 * L 包围框 + μ3 * L 伪检测框 Among them, μ1, μ2, and μ3 are preset weights of three loss functions.

6. A SAR image target detection system based on point features, characterized in that, It includes the following modules: A feature extraction module, which is used to input the SAR image into the feature extraction module to obtain feature maps of different scales; A point feature extraction module, which is used to perform convolutional processing on the feature map by inputting it into a point feature detection network, obtain the processed feature map, and output point features through a 1×1 convolution of the processed feature map; A pseudo detection box generation module, which is used to select the minimum and maximum x and y values in the point features as the coordinates of the upper left corner and the lower right corner of the pseudo detection box, and convert the point features into pseudo detection boxes; A predicted target bounding box module, which is used to perform deformable convolutional DCNv2 operations on the processed feature map and the point features, and then output a corrected box for the pseudo detection box and the target category through a 1×1 convolution of the operation result; Add the corrected box of the pseudo detection box and the pseudo detection box to obtain the final predicted target bounding box; Construct a loss function module for constructing a total loss function \(L\) based on the predicted class loss \(L_{cls}\), the predicted target bounding box loss \(L_{box}\), and the pseudo-detection box loss \(L_{pseudo}\), so as to train the point feature detection network, update the network parameters through backpropagation, repeat this step until the performance of the point feature detection network converges, save the network structure and parameters, and obtain a trained point feature detection network; 类别 、 the predicted target bounding box loss \(L_{box}\) 包围框 and the pseudo-detection box loss \(L_{pseudo}\) 伪检测框 Construct a total loss function \(L\) to train the point feature detection network and update network parameters through backpropagation. Repeat this step until the performance of the point feature detection network converges. Save the network structure and parameters to obtain a trained point feature detection network; A detection module, which is used to input the SAR image to be detected into the trained point feature detection network to obtain the target bounding box and target category information as the detection result.

Images