CN110910413A - ISAR image segmentation method based on U-Net - Google Patents
ISAR image segmentation method based on U-Net Download PDFInfo
- Publication number
- CN110910413A CN110910413A CN201911192861.3A CN201911192861A CN110910413A CN 110910413 A CN110910413 A CN 110910413A CN 201911192861 A CN201911192861 A CN 201911192861A CN 110910413 A CN110910413 A CN 110910413A
- Authority
- CN
- China
- Prior art keywords
- image
- net
- isar image
- isar
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000003709 image segmentation Methods 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 64
- 230000011218 segmentation Effects 0.000 claims abstract description 27
- 238000012545 processing Methods 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000005070 sampling Methods 0.000 claims description 54
- 239000011159 matrix material Substances 0.000 claims description 26
- 238000010606 normalization Methods 0.000 claims description 17
- 230000008569 process Effects 0.000 claims description 17
- 230000000694 effects Effects 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000011426 transformation method Methods 0.000 claims description 4
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 4
- 238000000605 extraction Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 14
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000002592 echocardiography Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 230000000877 morphologic effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 230000003042 antagnostic effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001427 coherent effect Effects 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 238000009432 framing Methods 0.000 description 1
- 210000002364 input neuron Anatomy 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010845 search algorithm Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/90—Dynamic range modification of images or parts thereof
- G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
- G06T2207/10044—Radar image
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the field of radar signal processing, relates to an ISAR image segmentation method, in particular to an ISAR image segmentation method based on U-Net, and comprises the following steps: s1, preprocessing data; s2, generating a training set; s3, constructing a network structure of the U-Net and setting network parameters; s4, training U-Net; and S5, carrying out data preprocessing on the target ISAR image to be segmented according to S1, sending the preprocessed image into the U-Net trained in S4, and finally obtaining the segmentation result of the target ISAR image through a network output layer. Compared with the traditional method, the method has superiority in the aspect of feature extraction, can obtain deep semantic features according to the image by utilizing deep learning, and realizes more accurate segmentation of the ISAR image; compared with the FCN, the method has the advantages that parameters required to be trained are few, the network can be trained only through one-time training and few training samples, and training efficiency is high.
Description
Technical Field
The invention belongs to the field of radar signal processing, relates to an ISAR image segmentation method, and particularly relates to an ISAR image segmentation method based on U-Net.
Background
With the high importance of aerospace in various countries of the world and the rapid development of civil aerospace, more and more satellites are launched and lifted off, and the space becomes increasingly crowded and more competitive and antagonistic. How to effectively sense the space situation and further control the space becomes a big problem currently faced by each aerospace large country. Inverse Synthetic Aperture Radar (ISAR) is an all-weather and all-day means, has the advantages of high resolution and large amount of acquired information, and plays an important role in extracting spatial target characteristics. The information such as the appearance, the structure and the size of the target can be inverted through the obtained high-resolution ISAR image of the space target, and the target posture can be further estimated by combining the multi-frame image.
And performing inversion estimation on the structure scale and the motion attitude of the satellite target by using the ISAR image, and firstly extracting the ISAR image contour characteristics of the target, namely realizing the segmentation of the target in the image. Image segmentation is a processing technology for extracting a region of interest in a map, and the quality of the processing effect directly influences the cost of subsequent processing. At present, a great number of algorithms exist in the field of image segmentation, and the algorithms include an edge detection method based on Canny operators, a method based on boundary tracking, a method based on region growing, separating and aggregating and the like. In addition, Rafael c. and Beucher et al propose morphological watershed-based segmentation methods that are highly interesting because of their fast computation speed and their ability to accurately locate the edges of images. However, these conventional image segmentation methods are sensitive to noise, do not consider images and semantic information between images, and have a limited application range.
In recent years, the deep learning technology has been rapidly developed, and an Image Semantic Segmentation (ISS) method based on deep learning is also changing day by day. Compared with the traditional image segmentation, the ISS adds certain semantic information to the target or the foreground in the image on the basis of the ISS, can obtain the information which needs to be expressed by the image according to the texture, the scene and other high-level semantic features of the image, and has higher practical value. In 2014, Grishick et al proposed a Regional Convolutional Neural Network (RCNN) based on a Convolutional Neural Network (CNN). The RCNN combines the candidate region generated by the selective search algorithm with the visual feature generated by the CNN, and simultaneously completes two tasks of target detection and ISS. However, RCNN has the disadvantages of severe dependence on candidate regions, insufficient segmentation accuracy and complex processing procedure, so Grishick et al further proposes an improved algorithm Fast-RCNN, maps the candidate regions onto the convolution feature map of CNN, and generates a feature map of a fixed size for each candidate region through pooling of regions of interest, thereby increasing the speed of generating the candidate regions. The Fast-RCNN Network proposed by Ren et al adds a Region suggestion Network (RPN) on the basis of the Fast-RCNN Network, further improving the speed of generating high-quality candidate regions. However, these candidate-region-based methods require a large number of candidate regions to be generated, and only rectangular framing of the target region can be achieved, and end-to-end pixel-level image segmentation cannot be completed. Long et al designs a Full Convolutional Network (FCN) compatible with any size image for image semantic segmentation, uses Convolutional layer to replace the full connection layer in the conventional CNN, and uses cross-layer and bilinear interpolation methods to extend image-level classification to pixel-level classification. The FCN directly classifies each pixel in the image without generating a target candidate region, and the original image directly outputs a classification result after passing through an end-to-end model. However, there are three problems with such FCN networks: firstly, after the image is subjected to pooling operation, the resolution of the characteristic graph is continuously reduced, and partial pixel spatial positions are lost; secondly, image context information cannot be considered in the segmentation process, and rich spatial position information cannot be fully utilized, so that the utilization rate of local features and global features is unbalanced; thirdly, the network structure is complex, and training needs a large amount of manually labeled samples.
Most of the methods are based on optical image segmentation, less researches are carried out on ISAR image segmentation reflecting the electromagnetic scattering characteristics of the target, and the problems of coherent speckle noise, side lobe interference, small sample number and the like which are specific to radar images are not considered.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides an ISAR image segmentation method based on U-Net. The U-Net adopts a network structure of an encoder-decoder, firstly carries out down sampling, learns the deep characteristics of an image through convolution of different degrees, and restores the deep characteristics of the image into the size of an input image through up sampling, and the up sampling is realized by deconvolution. The network structure of the U-Net only comprises a convolution layer and a pooling layer, and does not comprise a full connection layer, a shallow high-resolution layer in the network is used for solving the problem of pixel positioning, and a deep layer is used for solving the problem of pixel classification, so that segmentation at a semantic level can be realized, and speckle noise and side lobe influence of an ISAR image are eliminated. The method can fully utilize the spatial position information of the image without generating a candidate region, thereby greatly reducing the image segmentation time and improving the segmentation precision of the target ISAR image. Meanwhile, the training of the network can be completed by using a small number of marked ISAR images due to the simple network structure and few parameters.
In order to achieve the purpose, the invention is concretely realized by the following technical scheme:
s1 data preprocessing
S1.1 ISAR image normalization. Due to the comprehensive influence of the target environment, radar system parameters and target scattering characteristics, the target ISAR images obtained through imaging processing have different amplitudes, and in order to eliminate the influence of the image amplitude on image segmentation, the ISAR images composed of N matrix elements need to be subjected to amplitude normalization, wherein the normalization method comprises the following steps:
wherein s is an ISAR image data matrix, x is an ISAR image data matrix after amplitude normalization, n represents the nth element in the ISAR image data matrix, and the value of each element in the normalized matrix is in the range of 0 to 255.
S1.2, image enhancement. Due to the influence of the velocity compensation error and the side lobe, the ISAR image has object ghosting and bright lines, and the image quality is poor. In addition to the difference in the scattering intensity of echoes from different parts of the target, the "bright" parts in the ISAR image will occlude the "dark" parts. In order to facilitate image segmentation, the ISAR image needs to be enhanced, sharpened and better highlighted. Aiming at the characteristics of the ISAR image, a piecewise linear transformation method is adopted to increase the dynamic display range of the gray value of the image.
And S1.3, cutting the ISAR image subjected to image enhancement in the S1.2, and storing the ISAR image in a picture format. In general, a certain margin is left in the sampling bandwidth of the radar compared with the bandwidth of a transmitted signal, so that certain data redundancy exists in the ISAR image in the distance dimension. Meanwhile, the relatively stable motion state of the space target and the high pulse repetition frequency of the radar enable the ISAR image to have certain data redundancy in the azimuth dimension. The redundant data can increase the training time of the neural network and occupy the storage space of the computer, so that the ISAR image needs to be intercepted according to the actual size of the target, and the redundancy is eliminated. Meanwhile, in order to facilitate the further processing of the later U-Net, the intercepted ISAR image is stored into a picture format.
And S2, generating a training set. And taking the cut ISAR image sample in the S1.3 as a data set, and simultaneously carrying out manual marking on each image in the data set to obtain a label set corresponding to the data set, wherein the data set and the label set corresponding to the data set form a training set. Each sample in the training set contains an ISAR image and its artificially labeled ISAR image, and the training set is used as input for network training of U-Net.
S3, constructing a network structure of the U-Net and setting network parameters: U-Net adopts a network structure of an encoder-decoder, and comprises a down-sampling layer and an up-sampling layer, wherein the down-sampling layer consists of a convolution module, and the up-sampling layer consists of a deconvolution module. The upper sampling layer and the lower sampling layer adopt convolution operation with the same number of layers, and the SkipConnection structure is used for connecting the lower sampling layer and the upper sampling layer, and the specific implementation process is as follows:
s3.1, building a down-sampling layer of U-Net: the down-sampling layer consists of a plurality of (e.g., 4) convolution modules, each convolution module consisting of two 3 × 3 convolution layers, one modified Linear Unit (ReLU), and one 2 × 2 max pooling layer;
s3.2, building an upper sampling layer of the U-Net: the upsampling layer consists of a plurality of (e.g., 4) deconvolution modules, each of which consists of one 2 × 2 deconvolution (Up-Convolution) layer, two 3 × 3 Convolution layers, and one ReLU;
s3.3 Skip Connection: connecting the characteristic layer output by the convolution module in the lower sampling layer to the deconvolution module of the corresponding upper sampling layer, and connecting the characteristic layer and the deconvolution module in series to serve as input characteristics;
s3.4, constructing a Dropout layer: in order to avoid overfitting in the network training process, a Dropout layer is added into the lower sampling layer;
s3.5, a network output module of the U-Net is built: and (3) passing the output of the up-sampling layer through a convolution layer of 2 x 2 and 1 x 1 to obtain the final output of U-Net. Therefore, the U-Net can carry out end-to-end segmentation on the pixels, namely, an image is input, and the output is also an image with the same size.
S3.6, setting network parameters of U-Net: the network parameter setting of the U-Net comprises the number of convolution and deconvolution modules, an optimizer, a loss function, an activation function, Dropout and the like, and the setting of the network parameter is not uniformly specified and needs to be set according to the actual data type, the network application, the system requirement and other conditions.
And S4, training the U-Net. Training the U-Net constructed in the S3 by using an ISAR image training set generated in the S2 to obtain a trained network model; in order to ensure the training effect, part of data of a training set needs to be reserved as a verification set, and the network is subjected to performance evaluation in the training process to ensure the training effect.
And S5, carrying out data preprocessing on the target ISAR image to be segmented according to S1, sending the preprocessed image into the U-Net trained in S4, and finally obtaining the segmentation result of the target ISAR image through a network output layer.
The invention has the beneficial effects that:
(1) compared with the traditional method, the ISAR image segmentation method based on the U-Net has superiority in feature extraction, can obtain deep semantic features according to the image by utilizing deep learning, and realizes more accurate segmentation of the ISAR image.
(2) Unlike the method based on RCNN, the invention does not need candidate regions, cancels a full connection layer and only has a convolution layer and a pooling layer. The Skip Connection structure is adopted to connect the upper sampling layer and the lower sampling layer, so that the characteristics extracted by the lower sampling layer can be directly transmitted to the upper sampling layer, the pixel positioning of U-Net is more accurate, and the segmentation precision is higher.
(3) The method takes data as drive, can realize end-to-end segmentation, and simplifies the segmentation process of the target ISAR image; meanwhile, the method has good universality and strong generalization capability on ISAR images of different sizes, and meets the large-batch and real-time engineering requirements.
(4) The U-Net network has simple structure and shallow layer number, and compared with the FCN network, the U-Net network has less parameters to be trained, can complete the training of the network only by one-time training and less training samples, and has high training efficiency.
Drawings
FIG. 1 is a flow chart of an ISAR image segmentation method based on U-Net according to the present invention;
FIG. 2 shows an ISAR image of a satellite target after amplitude normalization;
FIG. 3 shows an ISAR image of a satellite target after image enhancement;
FIG. 4 shows a cropped ISAR image of a satellite target;
FIG. 5 is a schematic diagram of an ISAR image of a marked satellite target according to the present invention;
FIG. 6 is a schematic diagram of a U-Net network structure provided by the present invention;
FIG. 7 is a graph showing the variation of the loss function values with the number of iterations in the network training of the present invention;
FIG. 8 is a graph showing the variation of segmentation accuracy with iteration number in the network training of the present invention;
FIG. 9 shows the ISAR image segmentation result based on U-Net according to the present invention;
fig. 10 shows the result of ISAR image segmentation based on the conventional morphological method.
Detailed Description
The technical solution of the present invention is specifically described below with reference to the accompanying drawings and specific examples, and it should be noted that the technical solution of the present invention is not limited to the implementation manner described in the examples, and those skilled in the art refer to and refer to the content of the technical solution of the present invention, and the modifications and designs made on the basis of the present invention shall fall into the protection scope of the present invention.
The embodiment of the invention provides a U-Net-based ISAR image segmentation method, wherein the size of each ISAR image matrix is 1024 x 512, and the flow of the ISAR image segmentation method is shown in figure 1.
The method specifically comprises the following steps:
s1 data preprocessing
S1.1 ISAR image normalization. Due to the comprehensive influence of the target environment, radar system parameters and target scattering characteristics, the target ISAR images obtained through imaging processing have different amplitudes, and in order to eliminate the influence of the image amplitude on image segmentation, the ISAR images composed of N matrix elements need to be subjected to amplitude normalization, wherein the normalization method comprises the following steps:
wherein s is an ISAR image data matrix, x is an ISAR image data matrix after amplitude normalization, n represents the nth element in the image data matrix, and the value of each element in the normalized matrix is in the range of 0 to 255.
The normalization process of the ISAR image is also a graying process for converting the color image into a 256-step grayscale image, 0 representing black and 255 representing white. The gray-scale image only contains brightness information, can reflect the strength information of the electromagnetic scattering of the target, and meets the requirements of image segmentation. Meanwhile, because the color information is not contained, the image information amount is greatly reduced, the image processing calculation amount is correspondingly greatly reduced, and the subsequent processing is convenient. Fig. 2 shows the result of the normalization of the satellite target ISAR image.
S1.2, image enhancement. Due to the influence of the velocity compensation error and the side lobe, the ISAR image has object ghosting and bright lines, and the image quality is poor. In addition to the difference in the scattering intensity of echoes from different parts of the target, the "bright" parts in the ISAR image will occlude the "dark" parts. In order to facilitate image segmentation, the ISAR image needs to be enhanced, sharpened and better highlighted. Aiming at the characteristics of the ISAR image, a piecewise linear transformation method is adopted to increase the dynamic display range of the gray value of the image.
S1.2.1, setting a threshold value, and setting the element value which is larger than the threshold value in the ISAR image matrix after normalization in S1 as the size of the threshold value, wherein the element value is unchanged when the element value is smaller than the threshold value. In the embodiment, an Otsu threshold method is adopted to adaptively determine the threshold of each ISAR image obtained in S1.1, and the Otsu threshold method is an optimal algorithm selected for an image segmentation domain value, so that the calculation is simple and is not influenced by the image brightness and contrast.
S1.2.2 piecewise linear transformation. After threshold processing, contrast lifting is carried out on the image matrix within the threshold range, namely, piecewise linear transformation is adopted to transform the dynamic range of the image value range back to 0 to 255, which can be expressed as formula
Wherein x isnIs an element, r, in the amplitude normalized ISAR image data matrixnThresh is a threshold value for the elements of the ISAR image data matrix after the piecewise linear transformation, and n represents the nth element in the image data matrix. After contrast pull-up process, rnThe value of each element in (a) is still in the range of 0 to 255, corresponding to the range of grey values.
Fig. 3 shows an ISAR image of the satellite target after image enhancement.
And S1.3, cutting the ISAR image subjected to image enhancement in the S1.2, and storing the ISAR image in a picture format. In general, a certain margin is left in the sampling bandwidth of the radar compared with the bandwidth of a transmitted signal, so that certain data redundancy exists in the ISAR image in the distance dimension. Meanwhile, the relatively stable motion state of the space target and the high pulse repetition frequency of the radar enable the ISAR image to have certain data redundancy in the azimuth dimension. The redundant data can increase the training time of the neural network and occupy the storage space of the computer, so that the ISAR image needs to be intercepted according to the actual size of the target, and the redundancy is eliminated. Meanwhile, in order to facilitate the further processing of the later U-Net, the intercepted ISAR image is stored into a picture format.
In this embodiment, according to the size of the satellite target, the ISAR image is cut from the original 1024 × 512 size to 128 × 128 size, and the cut ISAR image is saved in a picture format by adopting a lossless compression PNG format. Fig. 4 shows the result of cropping the ISAR image of the satellite target.
And S2, generating a training set. And taking the cut ISAR image sample in the S1.3 as a data set, and simultaneously carrying out manual marking on each image in the data set to obtain a label set corresponding to the data set, wherein the data set and the label set corresponding to the data set form a training set. Each sample in the training set contains an ISAR image and its artificially labeled ISAR image, and the training set is used as input for network training of U-Net. The invention adopts labeling software Labelme opened by the Massachusetts institute of technology, carries out pixel-level labeling on the ISAR image stored in the PNG format by using the software, and simultaneously stores the ISAR image after labeling in the PNG format. Thus, the data set and the label set corresponding to the data set form a training set for network training of U-Net.
In the embodiment, a total of 50 satellite target ISAR images are marked, and fig. 5 shows the marking result of one satellite target ISAR image.
S3, constructing a network structure of the U-Net and setting network parameters: the U-Net adopts a network structure of an encoder-decoder, and comprises a down-sampling layer and an up-sampling layer, wherein the down-sampling layer consists of a convolution module, and the up-sampling consists of a back-convolution module. The upper sampling layer and the lower sampling layer adopt convolution operation with the same number of layers, and the SkipConnection structure is used for connecting the lower sampling layer and the upper sampling layer, and the specific implementation process is as follows:
s3.1, building a down-sampling layer of the U-Net; the downsampled layer consists of a plurality of (e.g., 4) convolution modules, each convolution module consisting of two 3 x 3 convolutional layers, one ReLU, and one 2 x 2 max pooling layer;
s3.2, building an upper sampling layer of the U-Net; the upsampling layer consists of a plurality of (e.g., 4) deconvolution modules, each of which consists of one 2 × 2 deconvolution (Up-Convolution) layer, two 3 × 3 Convolution layers, and one ReLU;
s3.3 Skip Connection; connecting the characteristic layer output by the convolution module in the lower sampling layer to the deconvolution module of the corresponding upper sampling layer, and connecting the characteristic layer and the deconvolution module in series to serve as input characteristics;
s3.4, constructing a Dropout layer: in order to avoid overfitting in the network training process, a Dropout layer is added into the lower sampling layer;
s3.5, a network output module of the U-Net is built: and (3) passing the output of the up-sampling layer through a convolution layer of 2 x 2 and 1 x 1 to obtain the final output of U-Net. Therefore, the U-Net can carry out end-to-end segmentation on the pixels, namely, an image is input, and the output is also an image with the same size.
S3.6, setting network parameters of U-Net: the network parameter settings of U-Net include the number of convolution and deconvolution modules, optimizer, loss function, activation function, Dropout, etc. In this embodiment, the size of the ISAR picture in the data set is 128 × 128, so the downsampling layer of the U-Net network is set to 5 convolution modules, the feature dimensions after convolution processing are sequentially set to 16-32-64-128-256, the corresponding upsampling layer is composed of 5 deconvolution modules, the output feature dimensions are sequentially set to 128-64-32-16, and the two layers form a symmetrical structure. Because the output of the output layer is an image, the activation functions of other layers are all ReLU functions except the Sigmoid function. The optimizer selects an Adam optimizer combining the advantages of two optimization algorithms of AdaGrad and RMSProp, so that the memory requirement is less, and the calculation is more efficient. To prevent the training process from overfitting, the Dropout layer is set to a 50% discard rate, i.e., the Dropout layer will randomly disconnect 50% of the input neurons each time the parameters are updated during the training process. And finally, selecting a binary cross entropy function as a loss function of the network.
Fig. 6 shows a schematic diagram of the U-Net network structure in this embodiment.
And S4, training the U-Net. Training the U-Net constructed in the S3 by using an ISAR image training set generated in the S2 to obtain a trained network model; in order to ensure the training effect, part of data of a training set needs to be reserved as a verification set, and the network is subjected to performance evaluation in the training process to ensure the training effect. In the embodiment of the invention, 20% of data is reserved as a verification set, and the verification set does not participate in training and is used for testing indexes of the model after each iteration is finished, such as loss functions, accuracy and the like. In this embodiment, the number of network iterations is set to 80.
Fig. 7 and fig. 8 respectively show the loss function value and the segmentation accuracy rate in the network training process as a function of the number of network training iterations. As can be seen from fig. 7 and 8, when the number of iterations reaches 60, the loss function values and the segmentation accuracy of the validation set and the training set both tend to be stable, which indicates that it is reasonable to set 80 iterations, and an effective training result can be obtained.
And S5, carrying out data preprocessing on the target ISAR image to be segmented according to S1, sending the preprocessed image into the U-Net trained in S4, and finally obtaining the segmentation result of the target ISAR image through a network output layer.
In order to verify the effect of the invention, the method of the invention is compared with a segmentation method based on traditional morphology, and the traditional method realizes the segmentation of the target ISAR image through corrosion, expansion and boundary extraction. Fig. 10 shows the result of the conventional method for segmenting the same satellite target ISAR image. Comparing fig. 9 and fig. 10, it can be seen that the method of the present invention is significantly superior to the conventional segmentation method, and can more accurately and reasonably extract the target contour without distortion of the contour edge.
From the results of the embodiment, it can be seen that the identification method of the present invention is significantly superior to the conventional segmentation method. Meanwhile, the network structure is simple, the required training samples are few, the requirement of real-time image segmentation can be met, and the method is suitable for small sample sets such as ISAR images.
Claims (6)
1. An ISAR image segmentation method based on U-Net is characterized by comprising the following steps:
s1 data preprocessing
S1.1 ISAR image normalization: in order to eliminate the influence of the image amplitude on image segmentation, the amplitude normalization of the ISAR image composed of N matrix elements is required, and the normalization method is as follows:
wherein s is an ISAR image data matrix, x is an ISAR image data matrix after amplitude normalization, n represents the nth element in the ISAR image data matrix, and the value of each element in the normalized matrix is in the range of 0 to 255;
s1.2, image enhancement: aiming at the characteristics of the ISAR image, a piecewise linear transformation method is adopted to increase the dynamic display range of the gray value of the image;
s1.3, cutting the ISAR image subjected to image enhancement in the S1.2, and storing the ISAR image in a picture format;
s2, generating a training set: taking the cut ISAR image sample in the S1.3 as a data set, and simultaneously carrying out manual marking on each image in the data set to obtain a label set corresponding to the data set, wherein the data set and the label set corresponding to the data set form a training set; each sample in the training set comprises an ISAR image and an artificially marked ISAR image thereof, and the training set is used as input for network training of U-Net;
s3, constructing a network structure of the U-Net and setting network parameters: the U-Net adopts a network structure of an encoder-decoder and comprises a down-sampling layer and an up-sampling layer, wherein the down-sampling layer consists of a convolution module, and the up-sampling layer consists of a deconvolution module; the upper sampling layer and the lower sampling layer adopt convolution operation with the same number of layers, and the SkipConnection structure is used for connecting the lower sampling layer and the upper sampling layer, and the specific implementation process is as follows:
s3.1, building a down-sampling layer of U-Net: the down-sampling layer consists of a plurality of convolution modules, and each convolution module consists of two convolution layers of 3 x 3, a correction linear unit and a maximum pooling layer of 2 x 2;
s3.2, building an upper sampling layer of the U-Net: the up-sampling layer consists of a plurality of deconvolution modules, and each deconvolution module consists of a 2 x 2 deconvolution layer, two 3 x 3 convolution layers and a ReLU;
s3.3 Skip Connection: connecting the characteristic layer output by the convolution module in the lower sampling layer to the deconvolution module of the corresponding upper sampling layer, and connecting the characteristic layer and the deconvolution module in series to serve as input characteristics;
s3.4, constructing a Dropout layer: in order to avoid overfitting in the network training process, a Dropout layer is added into the lower sampling layer;
s3.5, a network output module of the U-Net is built: the output of the up-sampling layer is processed by a 2 x 2 and 1 x 1 convolution layer to obtain the final output of U-Net, thereby ensuring that the U-Net can carry out end-to-end segmentation on pixels, namely, an image is input, and the output is also an image with the same size;
s3.6, setting network parameters of U-Net: the network parameter setting of the U-Net comprises the number of convolution and reverse convolution modules, an optimizer, a loss function, an activation function and Dropout, the setting of the network parameter has no uniform regulation and needs to be set according to the conditions of actual data types, network purposes, system requirements and the like;
s4, training U-Net: training the U-Net constructed in the S3 by using an ISAR image training set generated in the S2 to obtain a trained network model;
and S5, carrying out data preprocessing on the target ISAR image to be segmented according to S1, sending the preprocessed image into the U-Net trained in S4, and finally obtaining the segmentation result of the target ISAR image through a network output layer.
2. The U-Net based ISAR image segmentation method according to claim 1, wherein: in S4, in order to ensure the training effect, part of the data of the training set needs to be reserved as a verification set, and the network is subjected to performance evaluation in the training process to ensure the training effect.
3. The U-Net based ISAR image segmentation method according to claim 1, wherein: in S3.1, the downsampling layer consists of 4 convolution modules.
4. The U-Net based ISAR image segmentation method according to claim 1, wherein: in S3.2, the upsampling layer consists of 4 convolution modules.
5. The U-Net based ISAR image segmentation method according to claim 1, wherein: s1.2, the specific steps of image enhancement by adopting a piecewise linear transformation method are as follows:
s1.2.1 sets a threshold: setting the element value which is larger than the threshold value in the ISAR image matrix after the normalization of S1 as the threshold value, and keeping the element value which is smaller than the threshold value unchanged;
s1.2.2 piecewise linear transformation: after threshold processing, contrast lifting is carried out on the image matrix within the threshold range, namely, piecewise linear transformation is adopted to transform the dynamic range of the image value range back to 0 to 255, which can be expressed as formula
Wherein x isnIs an element, r, in the amplitude normalized ISAR image data matrixnAfter the contrast ratio is increased, r is the element of the ISAR image data matrix after the piecewise linear transformation, Thresh is a threshold value, n represents the nth element in the image data matrixnThe value of each element in (a) is still in the range of 0 to 255, corresponding to the range of grey values.
6. The U-Net based ISAR image segmentation method according to claim 5, wherein: in S1.2.1, the threshold value of each ISAR image obtained in S1.1 is adaptively determined by adopting an Otsu threshold value method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911192861.3A CN110910413A (en) | 2019-11-28 | 2019-11-28 | ISAR image segmentation method based on U-Net |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911192861.3A CN110910413A (en) | 2019-11-28 | 2019-11-28 | ISAR image segmentation method based on U-Net |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110910413A true CN110910413A (en) | 2020-03-24 |
Family
ID=69820235
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911192861.3A Pending CN110910413A (en) | 2019-11-28 | 2019-11-28 | ISAR image segmentation method based on U-Net |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110910413A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111967507A (en) * | 2020-07-31 | 2020-11-20 | 复旦大学 | Discrete cosine transform and U-Net based time sequence anomaly detection method |
CN112052762A (en) * | 2020-08-27 | 2020-12-08 | 西安电子科技大学 | Small sample ISAR image target identification method based on Gaussian prototype |
CN112132798A (en) * | 2020-09-18 | 2020-12-25 | 浙江大学 | Method for detecting complex background PCB mark point image based on Mini ARU-Net network |
CN112216293A (en) * | 2020-08-28 | 2021-01-12 | 北京捷通华声科技股份有限公司 | Tone conversion method and device |
CN113724238A (en) * | 2021-09-08 | 2021-11-30 | 佛山科学技术学院 | Ceramic tile color difference detection and classification method based on feature point neighborhood color analysis |
CN114966693A (en) * | 2022-07-20 | 2022-08-30 | 南京信息工程大学 | Airborne ship target ISAR refined imaging method based on deep learning |
CN115661144A (en) * | 2022-12-15 | 2023-01-31 | 湖南工商大学 | Self-adaptive medical image segmentation method based on deformable U-Net |
CN116703837A (en) * | 2023-05-24 | 2023-09-05 | 北京大学第三医院(北京大学第三临床医学院) | MRI image-based rotator cuff injury intelligent identification method and device |
CN117611828A (en) * | 2024-01-19 | 2024-02-27 | 云南烟叶复烤有限责任公司 | Non-smoke sundry detection method based on hyperspectral image segmentation technology |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107256396A (en) * | 2017-06-12 | 2017-10-17 | 电子科技大学 | Ship target ISAR characteristics of image learning methods based on convolutional neural networks |
US20190180443A1 (en) * | 2017-11-07 | 2019-06-13 | Align Technology, Inc. | Deep learning for tooth detection and evaluation |
CN110211137A (en) * | 2019-06-08 | 2019-09-06 | 西安电子科技大学 | Satellite Image Segmentation method based on residual error network and U-Net segmentation network |
CN110210422A (en) * | 2019-06-05 | 2019-09-06 | 哈尔滨工业大学 | It is a kind of based on optical imagery auxiliary naval vessel ISAR as recognition methods |
-
2019
- 2019-11-28 CN CN201911192861.3A patent/CN110910413A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107256396A (en) * | 2017-06-12 | 2017-10-17 | 电子科技大学 | Ship target ISAR characteristics of image learning methods based on convolutional neural networks |
US20190180443A1 (en) * | 2017-11-07 | 2019-06-13 | Align Technology, Inc. | Deep learning for tooth detection and evaluation |
CN110210422A (en) * | 2019-06-05 | 2019-09-06 | 哈尔滨工业大学 | It is a kind of based on optical imagery auxiliary naval vessel ISAR as recognition methods |
CN110211137A (en) * | 2019-06-08 | 2019-09-06 | 西安电子科技大学 | Satellite Image Segmentation method based on residual error network and U-Net segmentation network |
Non-Patent Citations (3)
Title |
---|
MANOHARA PAI M. M等: "Automatic Segmentation of River and Land in SAR Images: A Deep Learning Approach", 《2019 IEEE SECOND INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE AND KNOWLEDGE ENGINEERING (AIKE)》 * |
OLAF RONNEBERGER等: "U-Net: Convolutional Networks for Biomedical Image Segmentation", 《18TH INTERNATIONAL CONFERENCE ON MEDICAL IMAGE COMPUTING AND COMPUTER-ASSISTED INTERVENTION (MICCAI)》 * |
张富生: "基于卷积神经网络的小样本ISAR图像识别", 《电脑编程技巧与维护》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111967507A (en) * | 2020-07-31 | 2020-11-20 | 复旦大学 | Discrete cosine transform and U-Net based time sequence anomaly detection method |
CN112052762A (en) * | 2020-08-27 | 2020-12-08 | 西安电子科技大学 | Small sample ISAR image target identification method based on Gaussian prototype |
CN112216293A (en) * | 2020-08-28 | 2021-01-12 | 北京捷通华声科技股份有限公司 | Tone conversion method and device |
CN112132798A (en) * | 2020-09-18 | 2020-12-25 | 浙江大学 | Method for detecting complex background PCB mark point image based on Mini ARU-Net network |
CN112132798B (en) * | 2020-09-18 | 2022-04-29 | 浙江大学 | Method for detecting complex background PCB mark point image based on Mini ARU-Net network |
CN113724238B (en) * | 2021-09-08 | 2024-06-11 | 佛山科学技术学院 | Ceramic tile color difference detection and classification method based on feature point neighborhood color analysis |
CN113724238A (en) * | 2021-09-08 | 2021-11-30 | 佛山科学技术学院 | Ceramic tile color difference detection and classification method based on feature point neighborhood color analysis |
CN114966693A (en) * | 2022-07-20 | 2022-08-30 | 南京信息工程大学 | Airborne ship target ISAR refined imaging method based on deep learning |
CN114966693B (en) * | 2022-07-20 | 2022-11-04 | 南京信息工程大学 | Airborne ship target ISAR refined imaging method based on deep learning |
CN115661144A (en) * | 2022-12-15 | 2023-01-31 | 湖南工商大学 | Self-adaptive medical image segmentation method based on deformable U-Net |
CN116703837A (en) * | 2023-05-24 | 2023-09-05 | 北京大学第三医院(北京大学第三临床医学院) | MRI image-based rotator cuff injury intelligent identification method and device |
CN116703837B (en) * | 2023-05-24 | 2024-02-06 | 北京大学第三医院(北京大学第三临床医学院) | MRI image-based rotator cuff injury intelligent identification method and device |
CN117611828A (en) * | 2024-01-19 | 2024-02-27 | 云南烟叶复烤有限责任公司 | Non-smoke sundry detection method based on hyperspectral image segmentation technology |
CN117611828B (en) * | 2024-01-19 | 2024-05-24 | 云南烟叶复烤有限责任公司 | Non-smoke sundry detection method based on hyperspectral image segmentation technology |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110910413A (en) | ISAR image segmentation method based on U-Net | |
Jiang et al. | Edge-enhanced GAN for remote sensing image superresolution | |
CN113052210B (en) | Rapid low-light target detection method based on convolutional neural network | |
CN111222396B (en) | All-weather multispectral pedestrian detection method | |
CN112183258A (en) | Remote sensing image road segmentation method based on context information and attention mechanism | |
CN111626993A (en) | Image automatic detection counting method and system based on embedded FEFnet network | |
CN109741340B (en) | Ice cover radar image ice layer refined segmentation method based on FCN-ASPP network | |
CN112560717B (en) | Lane line detection method based on deep learning | |
CN113255837A (en) | Improved CenterNet network-based target detection method in industrial environment | |
CN113420794B (en) | Binaryzation Faster R-CNN citrus disease and pest identification method based on deep learning | |
CN114022408A (en) | Remote sensing image cloud detection method based on multi-scale convolution neural network | |
CN113052057A (en) | Traffic sign identification method based on improved convolutional neural network | |
CN116596792B (en) | Inland river foggy scene recovery method, system and equipment for intelligent ship | |
CN115830466A (en) | Glacier change remote sensing detection method based on deep twin neural network | |
CN114612306A (en) | Deep learning super-resolution method for crack detection | |
CN115047455A (en) | Lightweight SAR image ship target detection method | |
CN116363535A (en) | Ship detection method in unmanned aerial vehicle aerial image based on convolutional neural network | |
CN116258940A (en) | Small target detection method for multi-scale features and self-adaptive weights | |
CN113705340B (en) | Deep learning change detection method based on radar remote sensing data | |
CN114926826A (en) | Scene text detection system | |
CN116740572A (en) | Marine vessel target detection method and system based on improved YOLOX | |
CN117011160A (en) | Single image rain removing method based on dense circulation network convergence attention mechanism | |
CN116310892A (en) | Marine rescue method based on improved YOLOV4 target detection | |
CN115035429A (en) | Aerial photography target detection method based on composite backbone network and multiple measuring heads | |
CN114693712A (en) | Dark vision and low-illumination image edge detection method based on deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200324 |
|
RJ01 | Rejection of invention patent application after publication |