1. A low-frequency ultra-wideband synthetic aperture radar image leaf cluster covered target detection method comprises the following steps:

s1: generating a corresponding data set according to the application scene;

s2: constructing a neural network;

s3: training the neural network in step S2 using the data set in step S1;

s4: preliminarily predicting the target position by using the trained neural network;

s5: the output result of step S4 is processed to obtain a final prediction result.

2. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 1, wherein the specific process of the step S1 is as follows:

the method comprises the steps of collecting a plurality of images of a low-frequency ultra-wideband synthetic aperture radar leaf cluster covered target, wherein the images comprise single targets or multiple targets and information of the position of the target, generating a data set and a data label through random image cutting and data amplification based on a Gaussian probability model, wherein the data set is a plurality of image blocks, and the data label is a binary image correspondingly displaying the position of the target.

3. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 2, wherein the specific process of the step S2 is as follows:

and constructing a neural network comprising an 11-layer network structure in a Pythrch environment, wherein the neural network comprises 5 interleaved convolutional layers, 5 maximum pooling layers and a last 1 fully-connected layer, and a ReLU function is added after each 1 convolutional layer and each fully-connected layer to serve as an activation function.

4. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 3, wherein the specific process of the step S3 is as follows:

setting a learning rate, a parameter optimization function and a loss function by using the data set and the data label generated in the step S1, and training the network; through parameter debugging, training is stopped when the loss value of the network does not drop after a plurality of generations of training, and meanwhile, various regularization methods are selected to prevent overfitting of the network.

5. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 4, wherein the specific process of the step S4 is as follows:

firstly, an image to be detected is divided into a plurality of small areas and placed into a network for prediction, then splicing recovery is carried out, the areas are divided into partial edge overlapping areas, and because the result of the middle area output by the network is often more accurate than that of the edge area, the middle area output by the network is only taken as a preliminary prediction result, and the prediction effect of the network is improved.

6. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 5, wherein the specific process of the step S5 is as follows:

performing variance filtering on the network preliminary prediction result, namely filtering out an area with neighborhood variance smaller than a certain threshold value, and reserving an area with neighborhood variance larger than or equal to the threshold value; and after the variance filtering is carried out, carrying out binarization operation again to obtain a final predicted image.

7. The method for detecting the target covered by the leaf cluster of the low-frequency ultra-wideband synthetic aperture radar image according to claim 6, wherein the threshold is 0.02.

8. The method for detecting the target covered by the leaf cluster of the low-frequency ultra-wideband synthetic aperture radar image as claimed in claim 7, wherein in step S3, a mean square error MSE function is used as a loss function, a random gradient descent method with momentum is used as a network parameter optimization method, the momentum coefficient is 0.9, the batch processing size is 32, the initial learning rate is 0.1, and the initial learning rate decreases in power with the training algebra by taking 0.95 as a base number.

9. The method for detecting the target covered by the low-frequency ultra-wideband synthetic aperture radar image leaf cluster according to claim 8, wherein in the step S3, the network is trained for 80 generations.

10. The method for detecting the target covered by the leaf cluster of the low-frequency ultra-wideband synthetic aperture radar image as claimed in claim 8, wherein in step S2, the size of the network input image is set to 250 pixels x 250 pixels, and a normalization operation is performed; the output image size is 50 pixels by 50 pixels.

Background

In modern war, the importance of informatization reconnaissance means is increasingly highlighted. Under the stimulation of gulf wars and kosovor wars, a great deal of manpower and material resources are invested in western countries to research electronic reconnaissance technologies, and the detection of targets such as airplanes and vehicles by using electronic reconnaissance tools such as radars and the like on a battlefield is an indispensable means.

The area of the jungle in China is wide, on one hand, military facilities such as vehicles and the like of the party have good hidden environments, and on the other hand, the corresponding reconnaissance measures are inspired to be developed more by the party, and enemy vehicles and equipment are prevented from being hidden in the jungle. The low-frequency ultra-wideband synthetic aperture radar naturally becomes a main device for detecting the target covered by the leaf cluster by virtue of excellent penetrating performance.

The low-frequency ultra-wideband synthetic aperture radar image leaf cluster coverage target detection algorithm is developed under the background of deep learning, not only is the scientific technology of China chased the top level of the world under the stream of the information era, but also the national defense technology field is more likely to be the first place, thus being beneficial to enhancing the national defense construction and protecting the national security.

Current research efforts have focused primarily on the united states, for example, allen et al, usa scientific applications, propose a target detection method based on matched filter imaging; r.richard et al, the university of duck, usa, proposed a target detection method based on a hidden markov model; a target detection method based on spectral resolution is proposed by r.d. chaney et al in the lincoln laboratory; t.raju et al, Sandia laboratories, usa, proposed a target detection method based on spectral resolution; mitra et al, air force laboratories, usa, propose an order filter-based target detection method. Certainly, there are some domestic research achievements, which mainly focus on national defense science and technology university, for example, the goal detection method based on the median filter is researched by the doctor of national defense science and technology university; a target detection method based on clutter recognition is provided by a Fang academic doctor of the university of defense science and technology; and the Wang Guangzhou doctor of the university of defense science and technology proposes a change detection method applied to the detection of the hidden target of the leaf cluster.

The detection method is based on the traditional means of mathematical transformation, probability estimation, image filtering and the like, and is rarely combined with deep learning. In consideration of the fact that deep learning is very colorful in various fields in recent years, researchers naturally expect that the detection speed and the detection precision can be improved through the strong learning capability of the deep learning in the field of low-frequency ultra-wideband synthetic aperture radar image detection. Because the low-frequency ultra-wideband synthetic aperture radar image is not as rich and colorful as the optical image, has clear outline, has a large amount of clutter, has limited data set scale, and a deep learning method which can adapt to the characteristics of the low-frequency ultra-wideband synthetic aperture radar image still needs to be developed.

Disclosure of Invention

The invention provides a method for detecting the target covered by the leaf cluster of the low-frequency ultra-wideband synthetic aperture radar image, which has higher detection precision and better robustness.

In order to achieve the technical effects, the technical scheme of the invention is as follows:

a low-frequency ultra-wideband synthetic aperture radar image leaf cluster covered target detection method comprises the following steps:

s1: generating a corresponding data set according to the application scene;

s2: constructing a neural network;

s3: training the neural network in step S2 using the data set in step S1;

s4: preliminarily predicting the target position by using the trained neural network;

s5: the output result of step S4 is processed to obtain a final prediction result.

Further, the specific process of step S1 is:

the method comprises the steps of collecting a plurality of images of a low-frequency ultra-wideband synthetic aperture radar leaf cluster covered target, wherein the images comprise single targets or multiple targets and information of the position of the target, generating a data set and a data label through random image cutting and data amplification based on a Gaussian probability model, wherein the data set is a plurality of image blocks, and the data label is a binary image correspondingly displaying the position of the target.

Further, the specific process of step S2 is:

and constructing a neural network comprising an 11-layer network structure in a Pythrch environment, wherein the neural network comprises 5 interleaved convolutional layers, 5 maximum pooling layers and a last 1 fully-connected layer, and a ReLU function is added after each 1 convolutional layer and each fully-connected layer to serve as an activation function.

Further, the specific process of step S3 is:

setting a learning rate, a parameter optimization function and a loss function by using the data set and the data label generated in the step S1, and training the network; through parameter debugging, training is stopped when the loss value of the network does not drop after a plurality of generations of training, and meanwhile, various regularization methods are selected to prevent overfitting of the network.

Further, the specific process of step S4 is:

firstly, an image to be detected is divided into a plurality of small areas and placed into a network for prediction, then splicing recovery is carried out, the areas are divided into partial edge overlapping areas, and because the result of the middle area output by the network is often more accurate than that of the edge area, the middle area output by the network is only taken as a preliminary prediction result, and the prediction effect of the network is improved.

Further, the specific process of step S5 is:

performing variance filtering on the network preliminary prediction result, namely filtering out an area with neighborhood variance smaller than a certain threshold value, and reserving an area with neighborhood variance larger than or equal to the threshold value; after the variance filtering, performing binarization operation again to obtain a final predicted image; the threshold is 0.02.

Further, in step S3, a mean square error MSE function is used as a loss function, a random gradient descent method with momentum is used as a network parameter optimization method, the momentum coefficient is 0.9, the batch processing size is 32, the initial learning rate is 0.1, and the initial learning rate decreases in power with the training algebra with 0.95 as a base number; the network was trained for 80 generations.

Further, the network input image size is set to 250 pixels × 250 pixels, and a normalization operation is performed; the output image size is 50 pixels by 50 pixels.

Compared with the prior art, the technical scheme of the invention has the beneficial effects that:

the method comprises the steps of generating a data set of a low-frequency ultra-wideband synthetic aperture radar leaf cluster coverage target with certain representativeness and corresponding marks through a random cutting and data augmentation technology; designing a convolutional neural network for predicting a leaf cluster coverage target and outputting a predicted image with the same size as the input image; then training the designed neural network by using the generated data set, and stopping training when the network converges; and finally, designing a post-processing flow aiming at the network output image based on the characteristics of the network output result, wherein the post-processing flow comprises variance filtering and binarization operation, and a final prediction result is obtained. The scheme of the invention is suitable for detecting the low-frequency ultra-wideband synthetic aperture radar image leaf cluster covered targets, and can ensure that a plurality of targets are quickly detected under certain precision.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a diagram of a random rectangular intercept box generation implemented according to Gaussian probability distribution;

fig. 3 is a view showing an image obtained by rotating an original image by 4 degrees;

fig. 4 is a diagram showing an image obtained by increasing or decreasing the original image as a whole;

FIG. 5 is a diagram of a neural network architecture of the present invention;

FIG. 6 is a diagram showing the detection results of each stage of the image to be detected.

Detailed Description

The drawings are for illustrative purposes only and are not to be construed as limiting the patent;

for the purpose of better illustrating the embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product;

it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

As shown in fig. 1, a method for detecting a low-frequency ultra-wideband synthetic aperture radar image leaf cluster covered target includes the following steps:

s1: generating a corresponding data set according to the application scene;

s2: constructing a neural network;

s3: training the neural network in step S2 using the data set in step S1;

s4: preliminarily predicting the target position by using the trained neural network;

s5: the output result of step S4 is processed to obtain a final prediction result.

The specific process of step S1 is:

the method comprises the steps of collecting a plurality of images of a low-frequency ultra-wideband synthetic aperture radar leaf cluster covered target, wherein the images comprise single targets or multiple targets and information of the position of the target, generating a data set and a data label through random image cutting and data amplification based on a Gaussian probability model, wherein the data set is a plurality of image blocks, and the data label is a binary image correspondingly displaying the position of the target.

As shown in fig. 2, specifically, the present example is the detection of a leaf cluster covering target of a forest and a shrub in a certain area, and the image is a low-frequency ultra-wideband synthetic aperture radar image: for the detection of the low-frequency ultra-wideband synthetic aperture radar image leaf cluster covering target in forest and shrub, firstly, a data set is generated by using 2 images containing the target with 3000 pixels × 2000 pixels (1 pixel represents an area with 1m × 1 m). Taking 1 image as an example, the process of generating the data set comprises 2 steps, the first step is to cut, the center of the area where the target is located is taken as a Gaussian distribution center, the variance is set through an iteration method so that the target area is located in the area with the probability of the Gaussian distribution higher than 90%, and a rectangular frame is randomly generated to represent the cut area, so that a positive sample and a negative sample are obtained. The second step is to increase data, mainly by changing the image angle and changing the whole gray level of the image to generate image sets with diversity, to enhance the representativeness of the data sets, randomly selecting an angle alpha between 0-90 degrees for each image generated in the first step, then rotating the original image by alpha, alpha +90 degrees, alpha +180 degrees and alpha +270 degrees to obtain 4 rotated images, and adding the original image to obtain 5 images. Further, for these 5 images, the total gradation was increased by 10 and decreased by 10, respectively, to obtain 15 images, and a new data set was constructed. With the above method, 3885 samples are finally generated as the training set, and 1115 samples are generated as the verification set. Fig. 3 shows an image obtained by rotating the original by 4 degrees, and fig. 4 shows an image obtained by increasing or decreasing the original as a whole. As shown in fig. 3, the leftmost image is the original image, and the adjacent 4 images on the right side are different by 90 °; as shown in fig. 4, the leftmost image is the original, the middle image is the original whose overall tone is decreased by 10, and the rightmost image is the original whose overall tone is increased by 10.

The specific process of step S2 is:

as shown in fig. 5, a neural network including an 11-layer network structure is constructed in a pitorch environment, and includes 5 convolutional layers and 5 maximum pooling layers which are interleaved, and a last 1 fully-connected layer, wherein a ReLU function is added after each 1 convolutional layer and fully-connected layer as an activation function, a network input image size is set to 250 pixels × 250 pixels, and a normalization operation is performed; the output image size is 50 pixels by 50 pixels.

The specific process of step S3 is:

setting a learning rate, a parameter optimization function and a loss function by using the data set and the data label generated in the step S1, and training the network; through parameter debugging, training is stopped when the loss value of the network does not drop after a plurality of generations of training, and meanwhile, various regularization methods are selected to prevent overfitting of the network.

Specifically, the network is trained using the data sets and data labels generated in step S1. During training, an MSE (mean square error) function is used as a loss function, a random gradient descent (SGD) method with momentum is used as a network parameter optimization method, and the momentum coefficient is 0.9. The batch size is 32, the initial learning rate is 0.1, and decreases in power with the number of training generations at the base of 0.95. After 80 generations of network training, the network gradually converges and stops training.

The specific process of step S4 is:

firstly, an image to be detected is divided into a plurality of small areas and placed into a network for prediction, then splicing recovery is carried out, the areas are divided into partial edge overlapping areas, and because the result of the middle area output by the network is often more accurate than that of the edge area, the middle area output by the network is only taken as a preliminary prediction result, and the prediction effect of the network is improved. The size of the image to be detected is also 3000 pixels × 2000 pixels (1 pixel represents a region of 1m × 1 m), which does not meet the requirement of network input size, so the image is firstly divided into a plurality of image blocks of 250 pixels × 250 pixels, so that the image blocks can be input into the network. Because the confidence coefficient of the central area of the network output image is high, most image blocks are divided to have edge overlapping, then the middle area is intercepted as a network prediction result after the output result is obtained through network calculation, and the prediction result of the network for the whole image is obtained through splicing again.

The specific process of step S5 is:

performing variance filtering on the network preliminary prediction result, namely filtering out an area with neighborhood variance smaller than a certain threshold value, and reserving an area with neighborhood variance larger than or equal to the threshold value; and after the variance filtering is carried out, carrying out binarization operation again to obtain a final predicted image.

Specifically, variance filtering is performed on the network prediction result, that is, an area with neighborhood variance smaller than a certain threshold is filtered, an area with neighborhood variance larger than or equal to a certain threshold is reserved, and the threshold is set to be 0.02. The feature that the variance of the target area is different from that of the noise area in the network output result is used.

After the variance filtering, performing binarization operation again to obtain a final predicted image, wherein a black area in the image represents that no target exists, and a white area represents the position of the target.

For example, fig. 6 shows the detection results of each stage of the image to be detected by using the method of the present invention, fig. 6(a) shows the original image of the image to be detected, fig. 6(b) shows the preliminary detection result of the trained network on the image to be detected, fig. 6(c) shows the result of variance filtering in fig. 6(b), and fig. 6(d) shows the result of binarization in fig. 6(c), that is, the final detection result, which shows that the detection result is accurate.

The same or similar reference numerals correspond to the same or similar parts;

the positional relationships depicted in the drawings are for illustrative purposes only and are not to be construed as limiting the present patent;

it should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.