KR20180138189A - A method and apparatus for detecting target, training method for target detection - Google Patents

Abstract

Disclosed are a method for detecting a target and a device thereof and a learning method for target detection. According to one embodiment, the method for detecting a target, comprises the steps of: receiving polarized light data; determining a scale-up factor based on the number of the polarized light data; performing input scaling on the polarized light data based on the scale-up factor; and detecting a target based on the input scaled polarized light data.

Description

TECHNICAL FIELD [0001] The present invention relates to a method and an apparatus for detecting a target and a learning method for detecting a target,

The following embodiments are directed to a method and apparatus for detecting a target based on deep running.

Synthetic Aperture Radar (SAR) is widely used to monitor enemies in the defense field, because it can be used regardless of the weather, day and night, and has excellent penetrating power. However, SAR images are difficult to interpret intuitively due to severe Speckle noise.

Therefore, there is a demand for algorithms for analyzing and analyzing SAR images. Recent deep-running techniques have been applied to various object recognition and recognition problems and have shown remarkable performance. Algorithms using deep learning techniques to identify SAR images have also been studied.

The SAR image consists of four types of polarization data, HH, HV, VH, and VV. However, getting all the polarization data is not easy due to the cost of the machine and signal processing. In addition, some polarization data may be lost due to defects in the machine and communication equipment in the display situation.

Conventional SAR image identification studies process a single piece of polarized data. Therefore, there is a need for a target detection technique using all the polarizations using correlation between polarizations.

Embodiments can provide a technique for detecting a target.

Embodiments may also provide techniques for learning a neural network to detect a target.

According to an embodiment of the present invention, there is provided a target detection method comprising the steps of: receiving polarization data; determining a scale-up factor based on the number of the polarization data; Performing input scaling on the data, and detecting a target based on the input scaled polarization data.

The polarization data may include at least one of HH (Horizontal transmit and receive), HV (Horizontal transmit and receive), VH (vertical transmit and receive) and VV .

The determining may comprise determining the scale-up factor to be in inverse proportion to the number of the polarization data.

Wherein the performing step comprises the steps of: inputting 0 to the polarization data absent in the HH, HV, VH and VV polarization data; and applying the scale-up factor to the polarization data existing in the HH, HV, VH and VV polarization data And < / RTI >

The detecting may include transfer learning of the neural network using the input scaled polaris data and detecting the target based on the neural network that is learned by the transition.

Wherein the transition learning comprises: generating a task by combining the input scaled polynomial data in different numbers; generating a task with a smaller number of input scaled polynomial data, And repeatedly learning the neural network.

The detecting may further include concatenating feature maps that have passed through at least one layer included in the neural network, and performing feature map scaling on the feature maps.

A target detection apparatus according to an embodiment of the present invention includes a receiver for receiving polarization data and a controller for performing input scaling on the polarization data to detect a target, An input scaler for determining an up factor and performing input scaling on the polarization data based on the scale-up factor, and a detector for detecting the target based on the input scaled polarization data.

The polarization data may include at least one of HH (Horizontal transmit and receive), HV (horizontal transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)

The input scaler may determine the scale-up factor to be in inverse proportion to the number of the polarization data.

The input scaler may input 0 to the polarization data absent in the HH, HV, VH and VV polarization data and multiply the polarization data present in the HH, HV, VH and VV polarization data by the scale-up factor .

The detector can use the input scaled polaris data to perform a transfer learning of the neural network and detect the target based on the neural network that has been traversed.

The detector generates a task by differently combining the number of input scaled polarized data and sequentially inputs a task having a small number of input scaled polynomial data to be combined from a task having a large number of input scaled polarized data to be combined The neural network can be repeatedly learned.

The detector may concatenate feature maps that have passed through at least one layer included in the neural network, and may perform feature map scaling on the feature maps.

The neural network learning method according to an embodiment includes the steps of receiving polarized light data, determining a scale up factor based on the number of the polarized light data, Performing input scaling, and learning the neural network based on input scaled polarization data.

The polarization data may include at least one of HH (Horizontal transmit and receive), HV (horizontal transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)

The determining may comprise determining the scale-up factor to be in inverse proportion to the number of the polarization data.

Wherein the performing step comprises the steps of: inputting 0 to the polarization data absent in the HH, HV, VH and VV polarization data; and applying the scale-up factor to the polarization data existing in the HH, HV, VH and VV polarization data And < / RTI >

Wherein the step of learning comprises: generating a task by combining the number of input scaled polarized data differently; generating a task with a small number of input scaled polynomial data combinations from a plurality of tasks with a combined input scaled polynomial data And sequentially repeating the input of the neural network.

The learning may further include concatenating feature maps that have passed through at least one layer included in the neural network, and performing feature map scaling on the feature maps.

1 shows a schematic block diagram of a target detection device according to an embodiment.
Fig. 2 shows a schematic block diagram of the controller shown in Fig.
Fig. 3 shows an example of a neural network used by the detector shown in Fig.
Fig. 4 shows an example of an operation in which the detector shown in Fig. 2 links feature maps. Fig.
5A shows an example of a feature map for four polarization data inputs when input scaling is not performed.
5B shows an example of a feature map for HH polarization data input when input scaling is not performed.
5C shows an example of a feature map for HV polarization data input when input scaling is not performed.
5D shows an example of a feature map for VH polarization data input when input scaling is not performed.
5E shows an example of a feature map for VV polarization data input when input scaling is not performed.
6A shows an example of a feature map for four polarization data inputs when input scaling is performed.
6B shows an example of a feature map for HH polarization data input when input scaling is performed.
6C shows an example of a feature map for HV polarization data input when input scaling is performed.
6D shows an example of a feature map for VH polarization data input when input scaling is performed.
6E shows an example of a feature map for VV polarization data input when input scaling is performed.
FIG. 7 shows an example of feature maps generated according to whether input scaling and transition learning are performed.

In the following, embodiments will be described in detail with reference to the accompanying drawings. However, various modifications may be made in the embodiments, and the scope of the patent application is not limited or limited by these embodiments. It is to be understood that all changes, equivalents, and alternatives to the embodiments are included in the scope of the right.

The terms used in the examples are used for descriptive purposes only and are not to be construed as limiting. The singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, the terms "comprises" or "having" and the like refer to the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, But do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

The terms first, second, or the like may be used to describe various elements, but the elements should not be limited by terms. The terms may be named for the purpose of distinguishing one element from another, for example, without departing from the scope of the right according to the concept of the embodiment, the first element being referred to as the second element, The second component may also be referred to as a first component.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this embodiment belongs. Terms such as those defined in commonly used dictionaries are to be interpreted as having a meaning consistent with the contextual meaning of the related art and are to be interpreted as either ideal or overly formal in the sense of the present application Do not.

In the following description of the present invention with reference to the accompanying drawings, the same components are denoted by the same reference numerals regardless of the reference numerals, and redundant explanations thereof will be omitted. In the following description of the embodiments, a detailed description of related arts will be omitted if it is determined that the gist of the embodiments may be unnecessarily blurred.

A module in this specification may mean hardware capable of performing the functions and operations according to the respective names described in this specification and may mean computer program codes capable of performing specific functions and operations , Or an electronic recording medium, e.g., a processor or a microprocessor, equipped with computer program code capable of performing certain functions and operations.

In other words, a module may mean a functional and / or structural combination of hardware for carrying out the technical idea of the present invention and / or software for driving the hardware.

1 shows a schematic block diagram of a target detection device according to an embodiment.

Referring to FIG. 1, a target detecting apparatus 10 may receive data and detect or recognize a target from received data. The data may include visual data. For example, the data may include images and polarization data.

The image may include a Synthetic Aperture Radar (SAR) image, an Electro-Optical (EO) image, and an IR (InfraRed) image.

The polarization data may include at least one of HH (horizontal transmit and receive), HV (vertical transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)

The target detection device 10 can be used for target detection in a military application. For example, the target detecting apparatus 10 may be used in a SAR (Synthetic Aperture Radar).

The target detection device 10 can detect the target from the polarization data. The target detecting apparatus 10 can effectively detect the target using some of the polarization data in the case where it is difficult to acquire all the polarization data. That is, the target detection apparatus 10 can effectively detect the target by utilizing only some of the polarization data among HH, HV, VH, and VV.

The target detecting apparatus 10 can interpret a plurality of polarized light data as multi-modal data.

In the existing target detection method, a target is detected by separately generating and learning a neural network for classifying the target according to the type of the polarized data, which may be inefficient in terms of time and cost.

The target detecting apparatus 10 can detect the target using a single neural network irrespective of the types of polarization data or the number of combinations. That is, the target detecting apparatus 10 can effectively detect a target by using one multi-sensor network even if the types of polarization data and the number of combinations thereof are different.

The target detecting apparatus 10 includes a receiver 100 and a controller 200. [

The receiver 100 may receive the polarization data. The receiver 100 can output the received polarization data to the controller 200. [

The controller 200 can perform input scaling on the polarization data to detect the target. The controller 200 can perform input scaling on the polarization data to learn the neural network.

The controller 200 can detect the target using a plurality of neural networks. The controller 200 can exclude background objects in the data using one network. The controller 200 can use the other neural network to exclude the non-target and recognize the target.

The controller 200 can detect a region of interest (ROI) and recognize the target using a neural network. The controller 200 can detect the ROI by pre-detecting similar targets and removing non-targets.

Fig. 2 shows a schematic block diagram of the controller shown in Fig. 1, and Fig. 3 shows an example of a neural network used by the detector shown in Fig.

Referring to FIGS. 2 and 3, the controller 200 may include an input scaler 210 and a detector 230.

The input scaler 210 may determine a scale-up factor based on the number of polarization data and perform input scaling on the polarization data based on the scale-up factor.

The input scaler 210 may perform data augmentation to learn the neural network with a small number of data. For example, the input scaler 210 may increase the polarization data.

The input scaler 210 may rotate the polarization data to increase the polarization data. The input scaler 210 can rotate the original polarized light data at an arbitrary angle in consideration of the relationship between the shadow of the target and the scattering point. For example, the input scaler 210 may rotate the original polarized light data by setting the threshold of the rotation angle to 15 degrees.

The input scaler 210 may augment the data by pose synthesis of the polarization data. The input scaler 210 may select a plurality of polarization data and perform pose synthesis using the median angle of the pose angle of each polarization data.

In addition, the input scaler 210 may increase the polarization data using a random attaching method. At this time, the input scaler 210 can generate data that is robust to translation by arbitrarily arranging the polarization data in a predetermined size.

The input scaler 210 may determine the scale-up factor to be inversely proportional to the number of polarization data. For example, the input scaler 210 may determine a scale-up factor as shown in Table 1.

Number of Polarization Data Scale adjustment
(scale adjustments) Multiplication factor
(multiplication factors) Scale-up factor
(scale-up factors) 1 (ex HH) 1/1 12 12 2 (ex HH, HV) 1/2 12 6 3 (ex HH, HV, VH) 1/3 12 4 4 (ex HH, HV, VH, VV) 1/4 12 3

The input scaler 210 may input 0 to the polarization data absent among the HH, HV, VH, and VV polarization data. The input scaler 210 may multiply the polarization data present in the HH, HV, VH and VV polarization data by the scale-up factor.

For example, if there is one polarized data based on 12 being the least common multiple of 1, 2, 3, 4, which is the number of combinations of polarized light data, the input scaler 210 is set to 12 times the input polarized data value, The data value input to the neural network can be maintained within a certain range by multiplying the number of polarized light data by two when the two polarized light data exist and by multiplying by four when the three polarized light data exist.

This allows the neural network to extract features that are robust to changes in the number of polarization data. The input scaler 210 may output the four pieces of polarization data generated through the input scaling to the detector 230.

The detector 230 may detect the target based on the input scaled polarization data. The detector 230 may learn the neural network based on input scaled polarized data. The detector 230 can learn four channels of the four polarized data as inputs to learn the neural network.

The neural network used by the detector 230 may include CNN (Convolutional Neural Network). For example, the detector 230 may detect a target using a neural network having the structure of FIG.

In the example of FIG. 3, the neural network may include five convolution layers, three max-pooling layers, and two fully connected layers. The structure of the neural network is not limited to the example shown in Fig. 3, and the network structure such as the number of layers and the size of a layer may be different depending on a learning object or an object to be detected.

The value of the characteristic map of the convolution layer using the input of four channels can be expressed by Equation (1).

Here, W _hh , W _hv , W _vh , and W _vv may each denote a 3D filter, and * may denote a 3D convolution operation. B may refer to a bias term.

Referring to Equation (1), the feature map may vary depending on the number of polarization data (i.e., modal). The performance of the neural network may deteriorate as the number of polarized data changes.

The size of the characteristic map value may vary depending on the number of polarization data so that the characteristic of the neural network may be greatly influenced by the number of polarization data. Therefore, it may be difficult to stably learn the neural network when the number of polarized data is changed.

Since the detector 230 always receives all four channel inputs from the input scaler 210, it can learn the neural network stably. Detector 230 may perform input scaling multiple times.

The detector 230 may use the input scaled polarization data to transfer learning the neural network. The detector 230 can detect the target based on the neural network that has been learned by the transition.

The detector 230 may generate the task by differently combining the number of input scaled polarization data. The detector 230 may perform the transition learning by first learning the neural network for an easy task and by learning the neural network for an increasingly difficult task.

An easy task can mean a task that has various polarization data, and a difficult task can mean a task that has a single polarization data.

The detector 230 can sequentially learn a task having a small number of input scaled polarized data in which the number of input scaled polarized data to be combined is combined from many tasks and repeat the learning of the neural network.

The detector 230 may provide a good initial value for a difficult task by first performing learning on an easy task. The detector 230 through which the detector 230 can improve the performance of the network.

In addition, the detector 230 can perform learning including data of an easy task when learning a neural network for a difficult task so that the neural network does not lose its ability in an easy task.

For example, the detector 230 may learn the neural network in the order shown in Table 2. [

Learning stage One 2 3 4 Number of Polarization Data 4 4 and 3 4, 3 and 2 4, 3, 2 and 1

In the example of Table 2, the detector 230 may perform learning for four tasks with polarization data four, and learning for four tasks with polarization data four and three tasks. Thereafter, the detector 230 performs learning on tasks having four, three, and two pieces of polarization data, and then performs learning on tasks having four, three, two, and one pieces of polarization data .

Fig. 4 shows an example of an operation in which the detector shown in Fig. 2 links feature maps. Fig.

Referring to FIG. 4, the detector 230 may merge feature maps that have passed through at least one layer included in the neural network. The merging of the feature maps may include a concatenation of feature maps. That is, the detector 230 may merge the feature maps in a manner other than the connection.

The detector 230 may concatenate feature maps that have passed through at least one layer included in the neural network. For example, the detector 230 may connect the output feature maps after passing the input data of the four channels once for each channel.

The fact that you connect feature maps can mean that you attach feature maps. For example, if each feature map before linking has 16 channels, then it can be 64 channels after linking.

The detector 230 may multiply the scaling factor by the output feature map for the polarization data that is present in the HH, HV, VH, and VV polarization data.

Number of Polarization Data Scale adjustment
(scale adjustments) Multiplication factor
(multiplication factors) Scale-up factor
(scale-up factors) 1 (ex HH) 1/1 12 12 2 (ex HH, HV) 1/2 12 6 3 (ex HH, HV, VH) 1/3 12 4 4 (ex HH, HV, VH, VV) 1/4 12 3

The detector 230 may determine the scale-up factor to be in inverse proportion to the number of polarization data as well as the input scaler 210. [ For example, the detector 230 may determine a scale-up factor as shown in Table 3.

The detector 230 may input 0 to the polarization data absent from the HH, HV, VH, and VV polarization data. The detector 230 may multiply the scale-up factor in Table 3 for a feature map corresponding to polarization data present in HH, HV, VH and VV polarization data.

For example, the detector 230 may perform feature map scaling based on 12, which is the least common multiple of 1, 2, 3, 4, which is the number of combinations of polarization data. When there is one polarization data, the detector 230 multiplies the value of the feature map corresponding to the input polarization data by 12 and the value of each of the corresponding feature maps by 6, The data values of the characteristic maps of the neural network can be maintained within a constant range by multiplying the respective values of the corresponding feature maps by four times.

The order of scaling and feature map connections can be changed. For example, the detector 230 may connect the feature map after scaling.

The location of scaling and connecting feature maps can be changed as needed. For example, the detector 230 may perform scaling and connection on a feature map that has passed through a second layer or a third layer of the neural network. Scaling and linking of feature maps can also be done simultaneously.

Also, the linking and scaling of feature maps can be performed a plurality of times.

The feature maps used for feature mapping and scaling may have non-linearity after passing through the activation function. For example, the activation function may be ReLU (Rectified Linear Unit). Detector 230 may utilize a high nonlinear relationship between polarization data by connecting and scaling feature maps.

FIG. 5A shows an example of a feature map for four polarization data inputs when input scaling is not performed, and FIG. 5B shows an example of a feature map for HH polarized data input when input scaling is not performed.

FIG. 5C shows an example of a feature map for HV polarized data input when input scaling is not performed, and FIG. 5D shows an example of a feature map for VH polarized data input when input scaling is not performed.

Fig. 5E shows an example of a feature map for VV polarization data input when input scaling is not performed, and Fig. 6A shows an example of a feature map for four polarization data inputs when input scaling is performed.

FIG. 6B shows an example of a feature map for HH polarized data input when input scaling is performed, and FIG. 6C shows an example of a feature map for HV polarized data input when input scaling is performed.

FIG. 6D shows an example of a feature map for VH polarized data input when input scaling is performed, and FIG. 6E shows an example of a feature map for VV polarized data input when input scaling is performed.

Referring to FIGS. 5A to 6E, the target recognition rate for whether input scaling and transition learning are performed can be expressed as shown in Table 4. Here, " ALL " may mean the case where all the polarized light data exists.

Polarized light HH HV VH VV ALL Recognition rate (%) for single-polarized data 1 channel input 96.71 97.69 98.90 98.29 99.39 Recognition rate for 4-channel input without input scaling (%) 92.08 94.15 93.42 92.45 94.03 Recognition rate (%) for input scaled 4 channel input 95.74 97.81 98.66 96.59 99.39 Recognition rate (%) for input scaled 4 channel input and transition learning performance 97.44 98.42 98.17 98.42 99.51

Single polarization data In the case of 1 channel input, it can mean the recognition rate result of 5 independent networks that have performed learning separately for each polarization data. In this case, the input terminal may be constituted by one channel. In this case, performance is high but may be inefficient in terms of memory and learning time.

The second can be the recognition rate of the network that input 4 channel input without performing input scaling. Third, it can mean the recognition rate of the network that input scaling and input 4 channel input. Fourth, it can mean the recognition rate of the network using 4 channel input with input scaling and performing the transition learning together.

It can be seen that the performance is significantly increased compared to the second case. That is, the target detecting apparatus 10 can detect the target efficiently in terms of the recognition rate, the memory, and the learning time in comparison with the existing target detecting method.

Table 5 shows the target recognition rates according to input scaling, feature map scaling, and transition learning.

Polarized light HH HV VH VV ALL Recognition rate (%) for single-polarized data 1 channel input 96.71 97.69 98.90 98.29 99.39 Recognition rate for 4-channel input without input scaling (%) 92.08 94.15 93.42 92.45 94.03 Recognition rate (%) for input scaled 4 channel input and transition learning performance 97.44 98.42 98.17 98.42 99.51 Recognition rate (%) for input scaled 4-channel input, feature map scaling, and transition learning performance 98.05 98.42 98.66 98.05 99.51

Referring to Table 5, it can be seen that the best performance is obtained when both input scaling, feature map scaling, and transition learning are performed.

Tables 6 and 7 show the scale invariant difference of the feature map scale. Table 6 shows NMD (Normalized Mean Difference) between feature maps when input scaling is performed, and Table 7 shows NMD when input scaling is not performed.

Here, the NMD between the characteristic maps A and B can be calculated based on the equation (2).

Feature map index One 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 NMD (4pol, HH) 0.0421 0.0541 0.0542 0.0403 0.054 0.054 0.0542 0.0501 0.0543 0.2222 0.034 0.0522 0.0538 0.0541 0.0506 0.6726 NMD (4pol, HV) 0.0494 0.0672 0.0672 0.0398 0.0673 0.0671 0.0673 0.051 0.0672 0.3848 0.0385 0.0445 0.0648 0.067 0.0506 One NMD (4pol, VH) 0.0568 0.0572 0.0573 0.0335 0.0575 0.0573 0.0573 0.102 0.0572 0.2389 0.0389 0.0406 0.0624 0.0572 0.1016 0.5936 NMD (4_pol, VV) 0.0445 0.0634 0.0634 0.0395 0.0634 0.0633 0.0634 0.0646 0.0634 0.4435 0.0421 0.0457 0.0623 0.0633 0.0642 One

Feature map index One 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 NMD (4pol, HH) 0.6116 One One 0.6063 0.6071 One 0.6072 One 0.1492 One 0.8532 0.9881 0.6082 0.6135 One NaN NMD (4pol, HV) 0.5793 0.999 One 0.5846 0.5837 One 0.584 One One 0.1152 0.7328 0.9678 0.585 0.5803 0.1055 NaN NMD (4pol, VH) 0.624 0.0182 One 0.6161 0.6175 One 0.6177 0.2901 One One 0.8399 0.9965 0.6178 0.628 One NaN NMD (4_pol, VV) 0.5872 0.9998 0.0784 0.5945 0.5933 0.087 0.5927 0.2932 One 0.107 0.799 0.9491 0.5907 0.5805 One NaN

The NaN value in the table means that all feature maps are dead feature maps, and 1 can mean that one of the feature maps is a dead feature map.

FIG. 7 shows an example of feature maps generated according to whether input scaling and transition learning are performed.

Referring to FIG. 7, it can be seen that each polarization data has a similar weight in constructing the feature map. That is, each polarization data may have a low correlation.

The target detection apparatus 10 can remarkably reduce the number of dead filters using input scaling, feature map scaling or transition learning. The target detection apparatus 10 can reduce the number of dead filters by correlating the polarization data. Further, the target detecting apparatus 10 can make the neural network utilize all the capacities.

The target detection apparatus 10 can generate a feature map having a similar dynamic range regardless of the number of polarization data through input scaling. The target detecting apparatus 10 can make use of the input of various polarizing data when constructing the characteristic map so that the neural network can use the correlation between the polarized data.

The target detection apparatus 10 can help the neural network to find an optimal point through input scaling. In addition, the target detection apparatus can help the neural network find the optimum point through the transition learning, and improve the detection performance and the performance of the neural network.

The target detecting apparatus 10 can learn a neural network by utilizing a higher level correlation between polarized data through characteristic map scaling.

The method according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer readable medium may include program instructions, data files, data structures, and the like, alone or in combination. Program instructions to be recorded on the medium may be those specially designed and constructed for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Magneto-optical media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

The software may include a computer program, code, instructions, or a combination of one or more of the foregoing, and may be configured to configure the processing device to operate as desired or to process it collectively or collectively Device can be commanded. The software and / or data may be in the form of any type of machine, component, physical device, virtual equipment, computer storage media, or device , Or may be permanently or temporarily embodied in a transmitted signal wave. The software may be distributed over a networked computer system and stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

Although the embodiments have been described with reference to the drawings, various technical modifications and variations may be applied to those skilled in the art. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (20)

Receiving polarized light data;
Determining a scale-up factor based on the number of the polarization data;
Performing input scaling on the polarization data based on the scale-up factor; And
Detecting the target based on input scaled polarization data;
And detecting the target.
The method according to claim 1,
The polarization data
And at least one of HH (Horizontal transmit and receive), HV (vertical transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)
A method of detecting a target.
The method according to claim 1,
Wherein the determining comprises:
Determining the scale-up factor to be in inverse proportion to the number of the polarization data
And detecting the target.
3. The method of claim 2,
Wherein the performing comprises:
Inputting 0 to the polarization data absent among the HH, HV, VH, and VV polarization data; And
Multiplying the polarization data present in the HH, HV, VH and VV polarization data by the scale-up factor
And detecting the target.
The method according to claim 1,
Wherein the detecting comprises:
Transfer learning the neural network using the input scaled polarized data; And
Detecting the target based on a transition learned neural network
And detecting the target.
6. The method of claim 5,
Wherein the transition learning comprises:
Generating a task by differently combining the number of input scaled polarized data; And
Sequentially inputting a task having a small number of input scaled polarized data to be combined from a task having a large number of input scaled polarized data to be combined, and repeatedly learning the neural network
And detecting the target.
6. The method of claim 5,
Wherein the detecting comprises:
Concatenating feature maps that have passed through at least one layer included in the neural network; And
Performing feature map scaling on the feature maps
Further comprising the steps of:
A receiver for receiving polarization data; And
A controller for performing input scaling on the polarization data to detect a target;
Lt; / RTI >
The controller comprising:
An input scaler that determines a scale-up factor based on the number of the polarization data and performs input scaling on the polarization data based on the scale-up factor; And
A detector for detecting a target based on input scaled polarized data
And the target detection device.
9. The method of claim 8,
The polarization data
And at least one of HH (Horizontal transmit and receive), HV (vertical transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)
Target detection device.
9. The method of claim 8,
The input scaler comprising:
The scale-up factor is determined so as to be in inverse proportion to the number of the polarization data
Target detection device.
10. The method of claim 9,
The input scaler comprising:
0 is input to the polarization data absent in the HH, HV, VH, and VV polarization data, and the polarization data existing in the HH, HV, VH, and VV polarization data is multiplied by the scale-
Target detection device.
9. The method of claim 8,
The detector comprises:
And performing a transfer learning on the neural network using the input scaled polaris data and detecting the target based on the neural network that has been learned by the transition
Target detection device.
13. The method of claim 12,
The detector comprises:
Wherein the tasks are generated by differently combining the number of input scaled polarized data and sequentially inputting a task having a small number of input scaled polynomial data pieces combined from a task having a large number of input scaled polynomial data to be combined, Iterative
Target detection device.
13. The method of claim 12,
The detector comprises:
Concatenating feature maps that have passed through at least one layer included in the neural network, and performing feature map scaling on the feature maps
Target detection device.
Receiving polarized light data;
Determining a scale-up factor based on the number of the polarization data;
Performing input scaling on the plurality of polarization data based on the scale-up factor; And
Learning the neural network based on input scaled polarized light data
/ RTI >
16. The method of claim 15,
The polarization data
And at least one of HH (Horizontal transmit and receive), HV (vertical transmit and receive), VH (vertical transmit and receive) and VV (vertical transmit and receive)
Neural network learning method.
16. The method of claim 15,
Wherein the determining comprises:
Determining the scale-up factor to be in inverse proportion to the number of the polarization data
/ RTI >
17. The method of claim 16,
Wherein the performing comprises:
Inputting 0 to the polarization data absent among the HH, HV, VH, and VV polarization data; And
Multiplying the polarization data present in the HH, HV, VH and VV polarization data by the scale-up factor
/ RTI >
16. The method of claim 15,
Wherein the learning step comprises:
Generating a task by differently combining the number of input scaled polarized data; And
Sequentially inputting a task having a small number of input scaled polarized data to be combined from a task having a large number of input scaled polarized data to be combined, and repeatedly learning the neural network
/ RTI >
20. The method of claim 19,
Wherein the learning step comprises:
Concatenating feature maps that have passed through at least one layer included in the neural network; And
Performing feature map scaling on the feature maps
The neural network learning method further comprising:

Images