KR102140805B1 - Neural network learning method and apparatus for object detection of satellite images - Google Patents

Abstract

The present invention provides a method and apparatus for learning a neural network for object detection of a satellite image. In the present invention, the method for learning the neural network to improve object detection performance using a weighted object candidate box and a metric learning may be provided. According to the present invention, the neural network learning method for object detection of the satellite image comprises the steps of: receiving an image including an object and extracting a feature map; generating a first object candidate box surrounding the object based on the feature map; generating a second object candidate box using a mask to which a predetermined weight is applied to a region extended N times (here, N is a real number greater than 1) around the first object candidate box; applying the second object candidate box to a region of interest (RoI) pooling and inputting the second object candidate box to a classification feature vector extractor; and inputting a classification feature vector outputted from the classification feature vector extractor into a metric learning network to learn.

Description

Neural network learning method and device for object identification of satellite images {NEURAL NETWORK LEARNING METHOD AND APPARATUS FOR OBJECT DETECTION OF SATELLITE IMAGES}

Embodiments relate to a neural network learning method and apparatus, and more particularly, to a method and apparatus for training a neural network for object identification.

Currently, in a tactical situation, a skilled reader manually analyzes optical satellite images to identify targets, such as enemy ships and aircraft. However, it is difficult to manually and rapidly analyze a large number of satellite images in a current situation in which it is predicted that the number of acquired satellite image information such as a number of ultra-small satellite operating systems will increase. Accordingly, there is an increasing demand to automate military surveillance and reconnaissance systems by developing a deep learning model capable of automatically identifying targets in satellite images.

On the other hand, in recent years, with the development of deep learning, object identification deep learning models have been widely applied in public, industrial, and military fields. In particular, the object identification deep learning model is expected to be highly useful in security fields that are manually observed and analyzed by humans, such as CCTV and satellite images. As the utilization of the object identification deep learning model increases in various fields as described above, there is an increasing demand to accurately identify the position of the object in the input image and classify it into detailed classification. However, when the size of the object in the image is relatively small, the performance of the deep learning model tends to decrease.

In the case of a satellite image in which an image is acquired at a high altitude, the size of the object to be identified is small, so information for distinguishing the object is relatively insufficient. Due to this, it is difficult to identify the position of the object, and it is also difficult to classify the object into detailed categories. Therefore, a new method for improving object identification performance is required.

Advance 1: KR 10-2015-0108577 Predecessor 2: KR 10-2019-0056009 Predecessor 3: KR 10-2019-0014908

One task according to embodiments is to provide a neural network learning method and apparatus for object identification of satellite images.

Another task according to embodiments is to provide an object identification method and apparatus using a neural network.

The problem to be solved is not limited to the above-described problems, and the problems not mentioned will be clearly understood by a person having ordinary knowledge in the technical field to which the present invention pertains from this specification and the accompanying drawings.

As a technical means for achieving the above-described technical problem, according to an aspect of the present disclosure, as a neural network learning method for object identification of a satellite image, an image containing an object is received and a feature map is received. Extracting, generating a first object candidate box surrounding the object based on the feature map, N times (where N is a real number greater than 1) centered on the first object candidate box, a predetermined area Generating the second object candidate box using the weighted mask, the second object candidate box is applied to the RoI pooling (Region of interest pooling) and input to the Classification Feature Vector Extractor The neural network learning method may be provided, including the step of inputting and learning the classification feature vector output from the classification feature vector extraction unit into a metric learning network.

Also, the weight may decrease as it progresses from the center of the mask to the outer region.

In addition, the second object candidate box may be generated with an N-fold extended region and a Hadamard product of the mask (where N is a real number greater than 1) around the first object candidate box.

Also, the height and width of the second object candidate box can be learned.

In addition, the second object candidate box may further include the step of being applied to RoI pooling and input to a classification network.

In addition, the metric learning network may learn that the distances of the classification feature vectors of objects belonging to the same class are close, and the distances of the classification feature vectors of objects belonging to different classes are increased.

In addition, the metric learning network may output metric loss and calculate a loss function using the metric loss.

According to another aspect, as an object identification method using a neural network, inputting an image into a neural network learned by a neural network learning method for object identification of a satellite image, and classification corresponding to the image through the learned neural network An object identification method using a neural network including outputting a result may be provided.

According to another aspect, a recording medium recording a program for executing a neural network learning method for object identification of a satellite image or an object identification method using a neural network in a computer may be provided.

According to another aspect, as a neural network learning apparatus, including a memory and a processor, the processor receives an image (Image), extracts a feature map (Feature Map), and surrounds the object based on the feature map A first object candidate box is generated, and a second object candidate box is generated by applying a predetermined weight to an area corresponding to N times the height and width of the first object candidate box (where N is a real number greater than 1), The second object candidate box is applied to the RoI pooling (Region of interest pooling), input to the Classification Feature Vector Extractor, and metrics of the Classification Feature Vector output from the Classification Feature Vector Extractor A neural network learning apparatus that is learned by inputting into a learning network may be provided.

According to another aspect, as a neural network learning apparatus, including a memory and a processor, the processor inputs an image into a neural network learned by a neural network learning method for object identification of a satellite image, and the learned neural network An object identification device using a neural network that outputs a classification result corresponding to an image may be provided.

Through a neural network learning method and apparatus for object identification of satellite images according to an embodiment, it is possible to contribute to automation of a military surveillance reconnaissance system and improvement of surveillance efficiency.

In addition, it is possible to learn by reflecting not only the object but also the surrounding information of the object. Since the surrounding information of the object is reflected, information for classifying the object may be relatively large even when the object size in the satellite image is small. Accordingly, object identification performance may be improved. As the object identification performance is improved, analytical ability and speed can be improved by supporting a manual observation and analysis task.

In addition, as the neural network for object identification is learned by metric learning, it is possible to improve the accuracy of identifying the type of object.

The effects are not limited to the above-described effects, and the effects not mentioned will be clearly understood by those skilled in the art from the present specification and the accompanying drawings.

1 is a view for explaining the architecture of a neural network according to an embodiment.
2 is a diagram for explaining a relationship between an input feature map and an output feature map in a neural network according to an embodiment.
3 is a block diagram showing a hardware configuration of a neural network device according to an embodiment.
4 is a diagram for describing a convolution operation of a neural network.
5 is a block diagram illustrating a structure of a neural network for object identification according to an embodiment.
6 is a view for explaining RoI pooling according to an embodiment.
7 is a view for explaining a mask.
8 is a view for explaining a mask according to an embodiment.
9 is a view for explaining the results of applying the mask.
10 is a diagram for explaining metric learning.
11 is a flowchart illustrating a neural network learning method according to an embodiment.

The terminology used in the present exemplary embodiments has been selected from general terms that are currently widely used as possible while considering functions in the present exemplary embodiments, but this may vary depending on the intention or precedent of a person skilled in the art or the appearance of a new technology. . In addition, in certain cases, some terms are arbitrarily selected, and in this case, their meanings will be described in detail in the description of the corresponding embodiment. Therefore, the terms used in the present embodiments should be defined based on the meanings of the terms and contents of the present embodiments, not simply the names of the terms.

In the description of the embodiments, when it is said that a part is connected to another part, this includes not only the case of being directly connected, but also the case of being electrically connected with another component in between. In addition, when a part includes a certain component, this means that other components may be further included instead of excluding other components unless otherwise specified. In addition, the term "... unit" described in the embodiments means a unit that processes at least one function or operation, which may be implemented in hardware or software, or a combination of hardware and software.

The terms "consisting of" or "comprising" as used in the embodiments should not be construed to include all of the various components or steps described in the specification, and some or all of them. It should be construed that the steps may not be included, or may further include additional components or steps.

The description of the following embodiments should not be construed as limiting the scope of rights, and those that can be easily inferred by those skilled in the art should be interpreted as belonging to the scope of rights of the embodiments. Hereinafter, exemplary embodiments for illustrative purposes will be described in detail with reference to the accompanying drawings.

In an embodiment, classification refers to an act of classifying a class of an object in an image given as input. For example, in the case of the MNIST data set, a total of 10 numbers from 0 to 9 are classified into each class, and the deep learning model trained on the input of one random number image is the type of the input image. It is sorted and printed out from 0 to 9. Localization means outputting location information about an object in an image and an object in the image. The bounding box is mainly used. The bounding box can be mixed with the object candidate box.

Object detection or object detection means that classification and localization are performed simultaneously. Depending on the learning purpose of the neural network, only a specific object can be detected, or multiple objects can be detected.

The neural network refers to a computational architecture that models a biological brain. Recently, as neural network technology has been developed, research into analyzing input data and extracting valid information by using the neural network in various types of electronic systems has been actively conducted. Devices that process neural networks require large amounts of computation on complex input data. Therefore, in order to analyze a large amount of input data in real time using a neural network and extract desired information, a technique capable of efficiently processing an operation related to the neural network is required.

1 is a view for explaining the architecture of a neural network according to an embodiment.

Referring to FIG. 1, the neural network 1 may be an architecture of a deep neural network (DNN) or an n-layers neural network. The DNN or n-layer neural network may correspond to Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Deep Belief Networks, Restricted Boltzman Machines, and the like. For example, the neural network 1 may be implemented as a convolutional neural network (CNN), but is not limited thereto. In FIG. 1, some convolutional layers are shown in the convolutional neural network corresponding to the example of the neural network 1, but the convolutional neural network includes a pooling layer and a pulley connected in addition to the illustrated convolutional layer. A fully connected) layer may be further included.

The neural network 1 may be implemented with an architecture having multiple layers including input data, feature maps and output. In the neural network 1, input data is subjected to a convolution operation with a filter called a kernel, and as a result, feature maps are output. At this time, the generated output feature maps are input feature maps, and convolution operation with the kernel is performed again, and new feature maps are output. As a result of such a convolution operation being repeatedly performed, a result of recognition of characteristics of input data through the neural network 1 may be finally output.

For example, when data having a size of 24x24 pixels is input to the neural network 1 of FIG. 1, the input data may be output as feature channels of 4 channels having a size of 20x20 through convolution operation with the kernel. Thereafter, the size of the 20x20 feature maps is reduced through iterative convolution with the kernel, and finally, features of 1x1 size can be output. The neural network 1 filters and outputs robust features that can represent the entire data from the input data by repeatedly performing convolution and subsampling (or pooling) operations on multiple layers. Recognition results of input data can be derived.

2 is a diagram for explaining a relationship between an input feature map and an output feature map in a neural network according to an embodiment.

Referring to FIG. 2, in a layer 3 of a neural network, the first feature map FM1 may correspond to an input feature map, and the second feature map FM2 may correspond to an output feature map. The feature map may mean a data set in which various characteristics of input data are expressed. The feature maps FM1 and FM2 may have elements of a 2D matrix or elements of a 3D matrix, and a pixel value may be defined for each element. The feature maps FM1 and FM2 have a width W (or column), height H (or row), and depth D. At this time, the depth D may correspond to the number of channels.

The convolution operation of the first feature map FM1 and the kernel weight map WM may be performed, and as a result, the second feature map FM2 may be generated. The weight map WM filters characteristics of the first feature map FM1 by performing a convolution operation with the first feature map FM1 with a weight defined in each element. The weight map WM shifts the first input feature map FM1 in a sliding window manner to perform convolution with windows (or tiles) of the first input feature map FM1. During each shift, each of the weights included in the weight map WM may be multiplied and added to each of the pixel values of the overlapped window in the first feature map FM1. As the first feature map FM1 and the weight map WM are convolved, one channel of the second feature map FM2 may be generated. In FIG. 2, a weight map (WM) for one kernel is illustrated, but in reality, weight maps of a plurality of kernels are respectively convolved with a first feature map (FM1), and a second feature map (FM2) of a plurality of channels is used. This can be generated. The second feature map FM2 may correspond to an input feature map of the next layer. For example, the second feature map FM2 may be an input feature map of a pooling (or subsampling) layer.

1 and 2 are shown only for the schematic architecture of the neural network 1 for convenience of explanation. However, the neural network 1 may be implemented with more or fewer layers, feature maps, kernels, and the like, as illustrated, and its sizes may also be variously modified. Anyone skilled in the art can understand.

3 is a block diagram showing a hardware configuration of a neural network device according to an embodiment.

The neural network apparatus 300 may be implemented with various types of devices such as a personal computer (PC), a server device, a mobile device, and an embedded device. As a specific example, voice recognition, image recognition, image classification, etc. using a neural network Smart phones, tablet devices, AR (Augmented Reality) devices, Internet of Things (IoT) devices, autonomous vehicles, robotics, medical devices, and the like, which are performed, but are not limited thereto. Further, the neural network device 300 may correspond to a dedicated hardware accelerator (HW accelerator) mounted on the above device, and the neural network device 300 may be a neural processing unit (NPU), a dedicated module for driving a neural network, It may be a hardware accelerator such as a TPU (Tensor Processing Unit), a Neural Engine, but is not limited thereto.

Referring to FIG. 3, the neural network device 300 includes a processor 310 and a memory 320. In the neural network device 300 illustrated in FIG. 3, only components related to the present exemplary embodiments are illustrated. Accordingly, it is apparent to those skilled in the art that the neural network device 300 may further include other general-purpose components in addition to the components shown in FIG. 3.

The processor 310 serves to control overall functions for executing the neural network device 300. For example, the processor 310 generally controls the neural network device 300 by executing programs stored in the memory 320 in the neural network device 300. The processor 310 may be implemented as a central processing unit (CPU), graphics processing unit (GPU), or application processor (AP) provided in the neural network device 300, but is not limited thereto.

The memory 320 is hardware that stores various data processed in the neural network device 300. For example, the memory 320 can store data processed by the neural network device 300 and data to be processed. have. In addition, the memory 320 may store applications, drivers, and the like to be driven by the neural network device 300. The memory 320 includes random access memory (RAM) such as dynamic random access memory (DRAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and CD- ROM, Blu-ray or other optical disk storage, hard disk drive (HDD), solid state drive (SSD), or flash memory.

The processor 310 reads/writes neural network data, for example image data, feature map data, kernel data, and the like, from the memory 320 and executes the neural network using the read/written data. do. When the neural network is executed, the processor 310 iteratively performs a convolution operation between the input feature map and the kernel to generate data about the output feature map. At this time, the computation amount of the convolution operation may be determined depending on various factors such as the number of channels of the input feature map, the number of channels of the kernel, the size of the input feature map, the size of the kernel, and the precision of the value. Unlike the neural network 1 shown in FIG. 1, an actual neural network driven by the neural network device 300 may be implemented with a more complex architecture. The methods described below may be performed by the processor 310 and the memory 320 of the neural network device 300.

4 is a diagram for describing a convolution operation of a neural network.

In the example of FIG. 4, it is assumed that the input feature map 410 is 6x6 size, the original kernel 420 is 3x3 size, and the output feature map 430 is 4x4 size. Can be implemented with feature maps and kernels. In addition, the values defined in the input feature map 410, the original kernel 420, and the output feature map 430 are all exemplary values, and the present embodiments are not limited thereto. Meanwhile, the original kernel 420 corresponds to the binary-weight kernel described above.

The original kernel 420 performs a convolution operation while sliding in the input feature map 410 in a window size of 3x3. The convolution operation sums all the values obtained by multiplying each pixel value of a certain window of the input feature map 410 and the weight of each element of the corresponding position in the original kernel 420, so that the output feature map 430 Refers to the operation of obtaining the value of each pixel. Specifically, the original kernel 420 first performs a convolution operation with the first window 411 of the input feature map 410. That is, each pixel value 1, 2, 3, 4, 5, 6, 7, 8, 9 of the first window 411 is weight -1, -1, +1 of each element of the original kernel 420, respectively. Multiplied by +1, -1, -1, -1, +1, +1 respectively, resulting in -1, -2, 3, 4, -5, -6, -7, 8, 9 . Next, 3, which is the result of adding all of the obtained values -1, -2, 3, 4, -5, -6, -7, 8, and 9, is calculated, and the pixels in 1 row and 1 column of the output feature map 430 The value 431 is determined as 3. Here, the pixel value 431 of the first row and the first column of the output feature map 430 corresponds to the first window 411. In the same way, the convolution operation between the second window 412 of the input feature map 410 and the original kernel 420 is performed, so that the pixel value 432 of 1 row 2 column of the output feature map 430 is -3. Is decided. Finally, the convolution operation between the 16th window 413, which is the last window of the input feature map 410, and the original kernel 420 is performed, so that the pixel value 533 of 4 rows and 4 columns of the output feature map 430 is − 13 is decided.

That is, the convolution operation between one input feature map 410 and one original kernel 420 is multiplied by the values of each element corresponding to each other in the input feature map 410 and the original kernel 420 and the sum of the multiplication results. It can be processed by repeatedly performing, and the output feature map 430 is generated as a result of the convolution operation.

According to an embodiment, an image may be input to a neural network learned by a neural network learning method for object identification of a satellite image, and a classification result corresponding to the image may be output through the learned neural network. Classification results may include the location and type of object. For example, the identification result of whether the object is a ship or an aircraft may be output. The method of learning the neural network for object identification will be described later in detail with reference to FIGS. 5 to 11.

5 is a block diagram illustrating a structure of a neural network for object identification according to an embodiment. Referring to FIG. 5, the neural network may receive an image containing an object and extract a feature map. An object candidate box surrounding the object may be generated based on the feature map. Object candidate boxes of different sizes may be generated by using a mask weighted in an N-fold extended region around the object candidate box (where N is a real number greater than 1). The object candidate boxes having different sizes may be second object candidate boxes. The second object candidate box may be applied to RoI pooling (Region of interest pooling) and input to a classification feature vector extractor. The classification feature vector output from the classification feature vector extraction unit may be input to a metric learning network to be learned.

According to an embodiment, the second object candidate box may be applied to RoI pooling and input to a classification network. By providing the surrounding information of the object, classification performance may be improved.

6 is a view for explaining RoI pooling according to an embodiment.

Referring to FIG. 6, the object of interest of the output feature map is detected by the object candidate box, and the size of the detected object candidate box has been applied to RoI pooling as it is identified. On the other hand, in the present invention, an area of a size corresponding to n times the object candidate box can be expanded to apply a radial weight. The object candidate box processed through this process may be applied to RoI pooling and then input into the classification network. As a result, the identification performance may be improved because the network has an effect of additionally providing not only an object but also information around the object. The classification network can output and learn the class score and location of the final object box.

According to one embodiment, the weight may decrease as it progresses from the center to the outer region. As the distance from the object increases, the importance of surrounding information related to the object may decrease. Therefore, it can be learned using information having higher importance in proximity to an object.

은 제1 영역(710)의 높이와 너비이다.

는 제2 영역(730)의 높이와 너비이다. 7 is a view for explaining a mask, and FIG. 8 is a view for explaining a mask according to an embodiment. Referring to FIG. 7, the mask may be a matrix. The mask may include a first region 710 and a second region 730. The first region 710 may mean an existing RoI pooling region. The existing RoI pooling area may correspond to the first object candidate box. The second region 730 may be a RoI pooling region according to an embodiment.

Is the height and width of the first region 710.

Is the height and width of the second region 730.

는 수학식1에 의해 계산될 수 있다. For example, a mask

Can be calculated by Equation 1.

[Equation 1]

는 x축, y축에 대해서

와 임의의 제1 영역내의 좌표간 최단 맨해튼 거리를 의미한다. (

)

는 가중치의 정도를 조절하는 파라미터를 의미한다. 도 7은

인 경우의 예시일 뿐 다른 실수가 적용될 수 있음은 물론이다. here,

Is for the x-axis and y-axis

And the shortest Manhattan distance between coordinates in any first area. (

)

Means a parameter that controls the degree of weighting. Figure 7

Of course, this is only an example, and other mistakes may be applied.

According to an embodiment, the weight may decrease as the mask progresses from the first area to the second area.

는

의 N배(N은 1보다 큰 실수) 일 수 있다. 뉴럴 네트워크는 물체 식별에 적절한 N값을 학습할 수 있다. 이에 따라, 최대의 물체 식별 효율을 갖도록 적절한 물체 주변 정보가 뉴럴 네트워크에 입력 될 수 있다. According to one embodiment, the height and width of the second mask area may be learned. The height and width of the second object candidate box generated by the mask can be learned.

The

Can be N times (N is a real number greater than 1). The neural network can learn N values suitable for object identification. Accordingly, appropriate object surrounding information may be input to the neural network to have maximum object identification efficiency.

9 is a view for explaining a second candidate box to which a mask is applied. Referring to FIG. 9, the second object candidate box may be generated by Hadamard Product. A second object candidate box may be generated by an N-fold extension of the first object candidate box (where N is a real number greater than 1) and the Adamar product of the mask. Adammar product is an operation that multiplies each component of two matrices of the same size.

10 is a diagram for explaining metric learning according to an embodiment. Referring to FIG. 10, Vector1 is a classification feature vector of Object1, which is Class A. Vector2 is a classification feature vector of Object2, which is Class A. Vector3 is a classification feature vector of Object3, which is Class B. Prior to metric learning, the distance between vector 1 and vector 2, which are the same sub-category, is greater than the distance between vectors 1 and 3, which are different sub-categories. After metric learning, the distance between vector 1 and vector 2, which are the same classification, becomes closer, and the distance between vector 3, which is another classification, becomes longer.

As such, the distance between the classification feature vectors of objects belonging to the same class may be close, and the classification feature vectors of objects belonging to different classes may be learned to be distant. Through this, it is possible to improve the identification performance by better classifying detailed classifications (for example, types of ships) having different similarities (for example, ships) but having different characteristics.

According to an embodiment, the metric learning network may further include outputting metric loss and calculating a loss function. Various loss functions such as CosLoss, ArcLoss, and Triplet Loss can be applied.

11 is a flowchart illustrating a neural network learning method according to an embodiment. Since the neural network learning method illustrated in FIG. 11 is related to the embodiments described in the above-described drawings, the contents described in the above drawings may be applied to the method of FIG. 11 even though the contents are omitted below.

Referring to FIG. 11, in step 1110, the neural network receives an image and a feature map is extracted. In step 1120, an object candidate box is generated based on the feature map. In step 1130, a second object candidate box is generated by applying a weight to a region larger than the first object candidate box. In step 1140, the second object candidate box is applied to RoI pooling and input to the classification feature vector extraction unit. In step 1150, learning is performed by performing metric learning using the classification feature vector.

As described above, although the present invention has been described by limited embodiments and drawings, the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains have various modifications and variations from these descriptions. It is possible. Therefore, the scope of the present invention should not be limited to the embodiments, and should be determined not only by the claims to be described later, but also by the claims and equivalents.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as program modules, being executed by a computer. Computer readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically include computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transport mechanism, and includes any information delivery media.

The above description of the present invention is for illustration only, and those skilled in the art to which the present invention pertains will be able to understand that it can be easily modified to other specific forms without changing the technical spirit or essential features of the present invention. . Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the above detailed description, and it should be interpreted that all changes or modified forms derived from the meaning and scope of the claims and equivalent concepts thereof are included in the scope of the present invention. do.

Claims (11)

As a neural network learning method for object identification of satellite images,
Receiving an image including an object and extracting a feature map;
Generating a first object candidate box surrounding the object based on the feature map;
Generating a second object candidate box using a mask to which a predetermined weight is applied to an area that is expanded N times (where N is a real number greater than 1) around the first object candidate box;
The second object candidate box is applied to RoI pooling (Region of interest pooling) and input to a classification feature vector extractor; And
A method of learning a neural network for object identification of a satellite image, comprising inputting and learning a classification feature vector output from the classification feature vector extraction unit into a metric learning network. According to claim 1,
The weight is reduced as it progresses from the center of the mask to the outer region, a neural network learning method for object identification of a satellite image. According to claim 1,
The second object candidate box is an object of a satellite image, which is generated by an N-times extended region around the first object candidate box (where N is a real number greater than 1) and the Hadamard Product of the mask. Neural network learning method for identification. According to claim 1,
A method of learning a neural network for object identification of a satellite image, in which height and width of the second object candidate box are learned. According to claim 1,
The second object candidate box is applied to RoI pooling, further comprising the step of inputting to a classification network, a neural network learning method for object identification of a satellite image. According to claim 1,
The metric learning network learns to identify objects in a satellite image, such that the distances of the classification feature vectors of objects belonging to the same class are close, and the distances of the classification feature vectors of objects belonging to different classes are increased. Neural network learning method for. According to claim 1,
The metric learning network outputs metric loss, and calculates a loss function using the metric loss, a neural network learning method for object identification of a satellite image. As an object identification method using a neural network,
Inputting an image into a neural network learned by the learning method according to any one of claims 1 to 7; And
And outputting a classification result corresponding to the image through the learned neural network. A recording medium recording a program for executing a method according to any one of claims 1 to 7 on a computer. As a neural network learning device,
Memory; And
Including a processor,
The processor,
The feature map is extracted by receiving an image, a first object candidate box surrounding the object is generated based on the feature map, and N times the height and width of the first object candidate box ( Here, N is a real object greater than 1) by applying a predetermined weight to a second object candidate box, and the second object candidate box is applied to RoI pooling (Region of interest pooling) to classify feature vectors A neural network learning apparatus input to a classification feature vector extractor and learned by inputting a classification feature vector output from the classification feature vector extractor into a metric learning network. An object identification device using neural networks,
Memory; And
Including a processor,
The processor,
Using the neural network, which inputs an image into a neural network learned by the learning method according to any one of claims 1 to 6, and outputs a classification result corresponding to the image through the learned neural network. Object identification device.

Images