KR102163573B1 - Apparatus and method on generating synthetic data for training real-time object detection system - Google Patents

Abstract

The present invention provides a database unit in which a plurality of background images and a plurality of object masks are stored, a background selection unit that selects and acquires a background image from the database unit, and selects at least one object mask from the database unit, and selects at least one object mask. By synthesizing at least one final object mask obtained from at least one transformed object mask on the background image selected by the object mask setting unit and the background selection unit that transforms by applying at least one of blurring, scaling, and rotation angle adjustment It is possible to provide an apparatus and method for generating synthesized data including a synthesized image generating unit that generates a synthesized image.

Description

Synthetic data generation device and method for learning real-time object detection system {APPARATUS AND METHOD ON GENERATING SYNTHETIC DATA FOR TRAINING REAL-TIME OBJECT DETECTION SYSTEM}

The present invention relates to an apparatus and method for generating synthesized data, and to an apparatus and method for generating synthesized data for learning a real-time object detection system.

Recently, there is a demand for an object detection technology for identifying an object in various application fields such as Augmented Reality (AR) or security system. In particular, AR technology seeks a high level of user interaction to provide a more realistic experience, and it is very important to recognize the surrounding environment in real time.

If it is possible to recognize which object is included in the current image, sufficient information on the surrounding environment can be provided. Therefore, in order to recognize the surrounding environment in real time, an object detection technology must also be performed in real time.

In recent years, research has been continued to dramatically improve the performance of an object detection system by using an artificial neural network learned by a deep learning technique in a real-time object detection technology. In order for an artificial neural network to exhibit reliable performance, it must be trained using a large number of training data acquired in various environments.

Some common databases provide multiple images for some types of objects, so that learning data can be obtained. However, if the object to be detected does not exist in the databases, a situation arises in which the learning data must be produced.

Creating learning data for object detection is a very difficult task in practice because it is necessary to manually set the borders for all target objects existing in the image, and because it usually requires more than tens of thousands of training data to learn an artificial neural network model. to be.

In addition, there is another problem to be solved in the case of learning a real-time object detection system.

This is visual degradation due to complex interactions between objects and environments. Visual degradation is generally caused by motion blur and occlusion.

1 shows an example of an image in which visual deterioration is induced.

In FIG. 1, (a) shows an image in which visual deterioration occurs due to motion blur, and (b) shows an image in which visual deterioration occurs due to occlusion.

As shown in (a), motion blur refers to a phenomenon in which an object appears blurred in an image according to the motion of an object (here, a baseball bat), which mainly occurs in low-quality cameras. As the contour of the object is deformed by the motion blur, even if the object detection system has already been learned for each individual object, it results in the object not being detected.

As shown in (b), occlusion refers to a phenomenon in which a part of an object (here, a knife) is obscured in an image by a distractor (here, a human hand), and the object detection system By making it impossible to discriminate, it results in not being able to detect the object regardless of whether the object is learned.

If the object detection system needs to use learning data reflecting various types of visual deterioration for each object in order to learn to detect objects despite visual deterioration, the amount of required learning data increases exponentially. Causes problems.

Therefore, even if the object detection system is efficiently fabricated, there arises a problem that the object detection device cannot reliably detect the object due to the lack of learning data.

Korean Patent Publication No. 10-2017-0037024 (published on March 23, 2017)

An object of the present invention is to provide an apparatus and method for generating synthetic data capable of generating a plurality of data for learning to learn a real-time object detection system.

Another object of the present invention is to provide an apparatus and method for generating synthetic data that enables an object detection system to easily detect an object even in visual deterioration such as motion blur and occlusion.

In order to achieve the above object, an apparatus for generating synthetic data according to an embodiment of the present invention includes: a database unit storing a plurality of background images and a plurality of object masks; A background selection unit that selects and acquires a background image from the database unit; An object mask setting unit that selects at least one object mask from the database unit and transforms the selected at least one object mask by applying at least one of blurring, scaling, and rotation angle adjustment; And a composite image generator configured to generate a composite image by synthesizing at least one final object mask obtained from at least one object mask transformed to the background image selected by the background selection unit. Includes.

The object mask setting unit may include an object mask selection unit for selecting at least one object mask from among a plurality of object masks stored in the database unit; A blurring unit for selectively blurring the selected at least one object mask according to a predetermined blurring probability value; And a trimming unit configured to output at least one object mask transformed by adjusting the size and rotation angle of the at least one object mask. It may include.

The blurring unit may blur the object mask by using a blur kernel that blurs the object mask according to a blur size and a blur direction angle.

The synthesized data generating apparatus acquires at least one obstacle mask according to a predetermined obstacle probability value, synthesizes the obtained obstacle mask with an object mask output from the object mask setting unit, and converts the final object mask to the synthesized image generating unit. An object composition unit that transmits; It may further include.

The object synthesizer may deform by adjusting the size and rotation angle of at least one obstacle mask, and synthesize the transformed obstacle mask with the object mask.

The composite image generator arranges and synthesizes at least one final object mask on the background image, and if the number of final object masks is a plurality, an IOU (intersection of union) between the arranged plurality of final object masks is allowed. The placement position may be varied to be less than the value, and verification labels for each of at least one final object mask synthesized in the synthesized image may be generated.

According to another embodiment of the present invention for achieving the above object, a method for generating composite data includes: selecting one background image and at least one object mask from among a plurality of background images and a plurality of object masks obtained in advance; Transforming the selected at least one object mask by applying at least one of blurring, scaling, and rotation angle adjustment; And generating a composite image by synthesizing at least one final object mask obtained from the transformed at least one object mask on the selected background image. Includes.

Accordingly, the apparatus and method for generating synthesized data according to an embodiment of the present invention generates learning data that enables the object detection system to search for objects in various environments by separately obtaining and synthesizing a plurality of backgrounds and a plurality of objects. can do. In addition, by applying visual changes such as blurring and adjusting the size and direction to the object and synthesizing it with the background, the object detection system can detect objects in various states. Learning data that enables reliable detection of objects can be created and provided.

1 shows an example of an image in which visual deterioration is induced.
2 shows a schematic structure of an apparatus for generating learning data according to an embodiment of the present invention.
3 is a view for explaining the operation of each component of the learning data generating apparatus of FIG.
4 is a diagram comparing differences between other synthesized data and whether alpha channel blurring is performed.
5 shows a method of generating learning data according to an embodiment of the present invention.
6 shows an example frame for testing a method of generating learning data according to an embodiment of the present invention.
7 shows an experiment result of an object detection performance of an object detection system learned using learning data according to an embodiment of the present invention.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the implementation of the present invention, reference should be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing a preferred embodiment of the present invention with reference to the accompanying drawings. However, the present invention may be implemented in various different forms, and is not limited to the described embodiments. In addition, in order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components unless specifically stated to the contrary. In addition, terms such as "... unit", "... group", "module", and "block" described in the specification mean units that process at least one function or operation, which is hardware, software, or hardware. And software.

FIG. 2 is a schematic structure of an apparatus for generating learning data according to an exemplary embodiment of the present invention, and FIG. 3 is a diagram for explaining the operation of each component of the apparatus for generating learning data of FIG. 2.

Referring to FIGS. 2 and 3, when the apparatus for generating learning data according to the present embodiment is described, the apparatus for generating learning data includes a database unit 100, a background selection unit 200, an object mask setting unit 300, and an obstacle mask. A setting unit 400, an object synthesis unit 500, and a synthesized image generation unit 600 are included.

The database unit 100 stores a background image and an object image, which are source images for generating a composite image for learning an object detection system. The database unit 100 may include a background database 110 for storing a background image and an object mask database 120 for storing an object image, respectively.

The background database 110 may collect and store various background images so that the composite data can be generated similarly to a complex real environment.

In addition, the object mask database 120 stores images of various objects that may be search targets. Here, the object image is an image pulled out according to the outline of the object, and functions as a kind of mask that is superimposed on the background image when the composite image is generated. That is, in this embodiment, the object image is used as an object mask.

The object image stored in the object mask database 120 may be collected in various ways, but as an example, it may also be obtained through a separate auxiliary segmentation neural network.

The auxiliary segmentation neural network is also a kind of artificial neural network and requires learning, so it is not efficient in many cases. However, in this case, the object image for learning the auxiliary segmentation neural network can be collected by acquiring the image containing the object in the designated simple background and removing the background using a simple image masking algorithm for the acquired image. The neural network can acquire various object images.

The background selection unit 200 selects a background image to be synthesized from among a plurality of background images stored in the database unit 100 and transmits the selected background image to the synthesis image generation unit 600. The background selection unit 200 may randomly select a background or may select a background in a predetermined order so that objects may be synthesized in a situation in which objects are arranged in various environments in the synthesized image.

The object mask setting unit 300 selects N object masks (where N is a natural number) to be combined into a composite image by being placed on the background image selected by the background selection unit 200 among a plurality of object masks stored in the database unit 100 Then, the selected N object masks are visually transformed by applying at least one of motion blur, size adjustment, and rotation to the object synthesis unit 500. That is, the object mask setting unit 300 selects and transforms N object masks to be synthesized, thereby giving the selected object mask a visual variation effect that may occur due to various interactions.

The object mask setting unit 300 may include an object mask selection unit 310, a blurring unit 320, and a trimming unit 330.

The object mask selection unit 310 selects N object masks to be synthesized from among a plurality of object masks stored in the database unit 100.

The blurring unit 320 adds a motion blur effect to the N object masks. The blurring unit 320 imposes a motion blur effect in a motion blur synthesis method. The blurring unit 320 may use a motion blur kernel that adjusts the blurring size (w _mb ) of the object mask and the blurring direction angle (θ _mb ) with respect to the x-axis direction. The motion blur kernel has a linear characteristic, and the blur size (w _mb ) is an average pixel through the blur direction angle (θ _mb ) with respect to the center of the motion blur kernel.

The blurring unit 320 may add a blurring effect to the applied object mask by using the motion blur kernel as shown in FIG. 3. In this case, the blurring unit 320 may apply a blurring effect to the alpha channel of the object mask for natural fusion with the background image.

In addition, the blurring unit 320 may selectively apply a blurring effect to the object mask. For example, the blurring unit 320 may apply the blurring effect to only some of the N object masks applied according to the predetermined blurring probability value and not apply the blurring effect to the remaining object masks. Also, for the object mask to which the blurring effect is applied, the blurring size (w _mb ) and the blurring direction angle (θ _mb ) may be differently applied.

Meanwhile, the trimming unit 330 receives the object mask from the blurring unit 320, adjusts the size, rotates, and outputs it.

The trimming unit 330 adjusts the size of the object mask according to the scaling parameter s _obj and adjusts the rotation angle of the object mask with respect to the x axis to the rotation parameter θ _{obj in} order to produce objects arranged in various situations. Adjust accordingly. Here, the scaling parameter s _obj may be set as a ratio to the size of the background image. For example, when the scaling parameter s _obj is set to 0.5, the object mask may be adjusted to be half the size of the background image. In this case, so that the object mask does not exceed the size of the background image, the scaling parameter s _obj may be set as a relative value with respect to the length of the long side of the background image.

In addition, the scaling parameter s _obj and the rotation parameter θ _obj may be set to different values for each of the applied object masks.

Meanwhile, the trimming unit 330 trims the bounding box of the object mask and outputs it to the object synthesis unit 500 in response to the size and shape of the resized and rotated object mask.

As shown in FIG. 3, the obstacle mask setting unit 400 generates a distractor mask that causes occlusion of an object and transmits it to the occlusion synthesis unit. In this case, the obstacle mask setting unit 400 may not generate and transmit an obstacle mask for all object masks, but may selectively generate an obstacle mask according to a predetermined obstacle probability value.

In some cases, the database unit 100 may further include an obstacle mask database for storing an obstacle mask, and the obstacle mask setting unit 400 may selectively receive the obstacle mask stored in the database unit 100. .

Similar to the object mask setting unit 400, the obstacle mask setting unit 400 may adjust and output the size and rotation angle of the obstacle mask. At this time, the size and rotation angle of the obstacle mask are adjusted by the obstacle scaling parameter (s _dis ) and the obstacle rotation parameter (θ _dis ), and the obstacle scaling parameter (s _dis ) and the obstacle rotation parameter so that the obstacle mask does not overlay the object mask. (θ _dis ) may be determined in consideration of the scaling parameter s _obj and the rotation parameter θ _obj .

In addition, in some cases, the obstacle mask setting unit 400 may apply a blurring effect to the obstacle mask. That is, the obstacle mask setting unit 400 may be configured similarly to the object mask setting unit 400.

The object synthesis unit 500 generates a final object mask by synthesizing a modified object mask applied from the object mask setting unit 400 and an obstacle mask applied from the obstacle mask setting unit 400. Here, the obstacle mask is overlapped on the object mask, but is synthesized so that the object mask is not completely covered by the obstacle mask as described above. FIG. 3 shows a final object mask in which a bag as an object mask and a human arm as an obstacle mask are combined as an example.

Here, for convenience of explanation, the obstacle mask setting unit 400 and the object combining unit 500 are separately illustrated, but the obstacle mask setting unit 400 may be included in the object combining unit 500.

The composite image generation unit 600 generates a composite image by overlapping the N final object masks transmitted from the object synthesis unit 500 on the background image transmitted from the background selection unit 200. That is, as shown in FIG. 3, the composite image generator 600 may arrange a plurality of final object masks on one background image.

In this case, the composite image generator 600 may adjust so that the final object mask disposed on the previous background image is not overlapped and disposed. To this end, the composite image generator 600 sequentially randomly arranges the N final object masks. And the IOU (intersection of union) between the previously placed final object mask and the currently placed final object mask is measured, and if the measured IOU is more than a predetermined tolerance (ε _iou ), the position of the final object mask currently placed is reset can do. That is, the position of the final object mask can be repeatedly changed until the IOU becomes less than the allowable value (ε _iou ).

The composite image generator 600 may generate a composite image for a situation in which various objects are arranged in various environments, and the generated composite image may be used as training data for learning an object detection system. In particular, the composite image generation unit 600 automatically generates a ground truth label for each object when the composite image is generated by directly positioning the final object mask on the background image.

As a result, the apparatus for generating synthesized data according to an embodiment of the present invention makes it possible to acquire a large number of images for various environments and conditions and various objects by using a small number of background images and object images. Therefore, in order to learn the object detection system, it is possible to easily acquire a large amount of learning data.

4 is a diagram comparing differences between other synthesized data and whether alpha channel blurring is performed.

In FIG. 4, (a) shows a case where the blurring unit 320 of the object mask setting unit 300 applies motion blur to the alpha channel of the object mask to generate a composite image, and (b) shows the case where the motion blur is applied to the alpha channel. This shows the case of creating a composite image without applying blurring.

The alpha channel is a channel for effectively fusing two colors when the color of one pixel in the graphic overlaps the color of another pixel. If motion blur is applied except for the alpha channel in the object mask, that is, for the alpha channel. When motion blur is not applied, as shown in the enlarged image of (b), there is a problem that artifacts are generated in the composite image. Therefore, in order to generate reliable training data, it is preferable that an alpha channel is also applied when applying a motion blur effect to an object mask.

5 shows a method of generating learning data according to an embodiment of the present invention.

Referring to FIGS. 2 and 3, the method of generating the training data of FIG. 5 will be described. First, the background selection unit 200 selects one background image from among a plurality of background images stored in the database unit 100, and an object mask The setting unit 300 selects N object masks from among a plurality of object masks (S10).

In addition, the object mask setting unit 300 determines whether to perform motion blur on each of the selected N object masks according to a supported blurring probability value (S20). If it is determined that motion blur is to be performed, the object mask setting unit 300 adds a motion blur effect by motion blur to the object mask according to the blur size (w _mb ) and the blur direction angle (θ _mb ). (S30). In this case, the blurring size (w _mb ) and the blurring direction angle (θ _mb ) may be adjusted differently for each object mask.

In addition, the object mask setting unit 300 adjusts the size of each of the N object masks according to the scaling parameter s _obj , and adjusts the rotation angle of the object mask with respect to the x axis according to the rotation parameter θ _obj (S40). ). In this case, the object mask setting unit 300 may adjust the size of the object mask so as not to exceed the size of the background image. In addition, the object mask setting unit 300 may trim the bounding box of the object mask in response to the size and shape of the resized and rotated object mask.

Further, the obstacle mask setting unit 300 determines whether to add an obstacle mask to the object mask according to a predetermined obstacle probability value (S50). When it is determined to add an obstacle mask, the obstacle mask setting unit 300 selects an obstacle to be added, and the object mask setting unit 300 sets the obstacle mask selected by the obstacle mask setting unit 300 to the object mask setting unit 300 ) To generate a final object mask by combining it with an object mask (S60).

In this case, the obstacle mask setting unit 300 may adjust the size and rotation angle of the obstacle mask similar to the object mask setting unit 400, and in some cases, a blurring effect may be applied.

The composite image generation unit 600 generates a composite image by overlapping the N final object masks transmitted from the object synthesis unit 500 with the background image selected by the background selection unit 200 (S70). The composite image generator 600 may determine whether the IOU is equal to or greater than a predetermined allowable value ε _iou , and adjust so that the IOU is not overlapped with the final object mask disposed on the previous background image. At this time, the composite image generator 600 automatically generates a ground truth label for each object in the composite image.

Hereinafter, results of testing the learning performance of learning data generated according to the learning data generation apparatus and method of the present embodiment are shown.

Here, as an example, as shown in Table 1, 12 object mask categories are selected, and the number of object masks for each category may be different.

When generating training data, the blurring size (w _mb ) was set to be selected from {20, 40}, and the blurring direction angle (θ _mb ) was set to be selected from {-45, 0, 45, 90}. And the scaling parameter (s _obj ) and the obstacle scaling parameter (s _dis ) are selected from {0.2, 0.3, 0.4}, and the rotation parameter (θ _obj ) and the obstacle rotation parameter (θ _dis ) are {-45, 0, 45, 90} was set to be selected. Also, the allowable value (ε _iou ) for IOU was set to 0.1.

Table 2 shows the object detection performance when the object detection system is trained using the synthesized image according to the embodiment of the present invention as training data, and the performance of the object detection system is evaluated using an average precision (AP) metric. .

Table 2 shows the results of testing by classifying the training data set into 4 types so that the performance against visual degradation can be compared when analyzing the learning performance of an object detection system using synthetic data.

Hereinafter, SynDB _n refers to the number of objects included in n (n is a natural number) images. In other words, SynDB ₁ in Table 2 means that one object is included in the image.

And w/none means a training data set that does not include motion blur and occlusion, w/occl means a training data set that includes occluded objects, and w/mb means a training data set that includes objects that use motion blur. And, w/mb and occl refer to a training data set that includes both motion blur and occlusion.

Referring to Table 2, it can be seen that the object detection system learned by using the training data generated according to the present embodiment exhibits very excellent performance with respect to an image in which visual deterioration has occurred.

6 shows an example image for testing training data generated according to an embodiment of the present invention.

In FIG. 6, (a) and (b) show examples of learning data obtained by the conventional method in order to compare and analyze the learning performance of the object detection system using synthetic data. I assume. And (c) shows an example of the training data generated according to the present embodiment, and (c) is a composite image in which a background image and an object mask are combined, so a verification label is automatically generated.

(a) is an image (N = 1) in which one of the objects of the object category is included in the image shown in Table 1, and the top image includes a gun, and the bottom image includes a broom. And (b) is an image including a plurality of objects among objects of the object category in the image. The upper image includes a bag, a bat, and a chair, and the lower image includes a bottle, knife, gun, and phone.

Meanwhile, (c) is a composite image in which a background image and an object mask are combined, and includes a plurality of objects.

Table 3 compares the results of learning the object detection system using the learning data obtained by the conventional method and the learning data generated according to the present embodiment as shown in FIG. 6.

Table 3 shows the results of testing while varying the number of objects included in the image from 1 to 8 (n = 1 to 8), the left is the actual image obtained by hand with the verification label, and the right is This refers to a synthesized image synthesized according to the present embodiment, and a case where visual deterioration of an object is included in all images was tested.

As shown in Table 3, it can be seen that the object detection system learned using the synthesized image generated by the learning data generation method according to the present embodiment exhibits better object detection performance regardless of the number of objects.

7 shows an experiment result of an object detection performance of an object detection system learned using learning data according to an embodiment of the present invention.

In FIG. 7, (a) shows the detection result of the occluded object, and (b) shows the detection result of the motion-blurred object. And in FIG. 7, the upper part shows the result of learning using a composite image that does not include motion blur and occlusion, the middle part shows the result of learning using the composite image including the occlusion, and the lower part shows the result of learning by using the composite image including motion blur and occlusion. It shows the result of training using all included composite images.

Referring to FIG. 7, the object detection system learned using a composite image that does not include motion blur and occlusion at the top detects only the area where occlusion does not appear in (a), whereas in (b) where motion blur exists, the object Was not detected at all.

In addition, the object detection system learned using the composite image containing the occlusion in the middle detects most areas of the object even though the occlusion occurred in (a), whereas in (b) where motion blur still exists, the object is completely detected. Could not be detected.

Finally, the object detection system learned using the composite image containing both motion blur and occlusion at the bottom can detect the object satisfactorily in both the occlusion (a) and the motion blur (b). Able to know.

As a result, the synthesized image generated by the synthesized data generating apparatus and method according to an embodiment of the present invention generates a synthesized image by inserting various objects into the background, so that the object detection system can easily detect objects in various environments. do. In addition, by reflecting visual deterioration phenomena such as motion blur and occlusion to each object included in the composite image, the object detection system can be trained to detect an object despite visual deterioration. Therefore, the performance of the object detection system can be greatly improved.

The method according to the present invention may be implemented as a computer program stored in a medium for execution on a computer. Here, the computer-readable medium may be any available medium that can be accessed by a computer, and may also include all computer storage 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, and ROM (Read Dedicated memory), RAM (random access memory), CD (compact disk)-ROM, DVD (digital video disk)-ROM, magnetic tape, floppy disk, optical data storage device, and the like.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely exemplary, and those of ordinary skill in the art will appreciate that various modifications and other equivalent embodiments are possible therefrom.

Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

100: database unit 200: background selection unit
300: object mask setting unit 400: obstacle mask setting unit
500: object synthesis unit 600: synthesis image generation unit

Claims (12)

A database unit storing a plurality of background images and a plurality of object masks;
A background selection unit that selects and acquires a background image from the database unit;
An object mask setting unit that selects at least one object mask from the database unit and transforms the selected at least one object mask by applying at least one of blurring, scaling, and rotation angle adjustment; And
A composite image generator configured to generate a composite image by synthesizing at least one final object mask obtained from at least one object mask transformed to the background image selected by the background selector; Including,
An object synthesis unit that obtains at least one obstacle mask according to a predetermined obstacle probability value, synthesizes the obtained obstacle mask with an object mask output from the object mask setting unit, and transmits the final object mask to the synthesis image generation unit; Contains more
The object composition unit
Deforming by adjusting the size and rotation angle of at least one obstacle mask, and synthesizing the transformed obstacle mask with the object mask,
The object combining unit synthesizes the obstacle mask to overlap the object mask to cover a part of the object mask, and applies a scaling parameter and a rotation parameter so that the obstacle mask does not overlay the object mask. The method of claim 1, wherein the object mask setting unit
An object mask selection unit selecting at least one object mask from among a plurality of object masks stored in the database unit;
A blurring unit for selectively blurring the selected at least one object mask according to a predetermined blurring probability value; And
A trimming unit configured to output at least one object mask modified by adjusting the size and rotation angle of the at least one object mask; Synthetic data generating device comprising a. The method of claim 2, wherein the blurring part
A composite data generation apparatus for blurring an object mask using a blur kernel that blurs the object mask according to a blur size and a blur direction angle. delete delete The method of claim 1, wherein the composite image generator
At least one final object mask is arranged and synthesized on the background image, and if the number of final object masks is multiple, the IOU (intersection of union) between the arranged multiple final object masks is arranged to be less than a predetermined allowable value. Variable position,
Synthetic data generation apparatus for generating a verification label for each of at least one final object mask synthesized in the synthesized image. Selecting one background image and at least one object mask from among a plurality of previously acquired background images and a plurality of object masks;
Transforming the selected at least one object mask by applying at least one of blurring, scaling, and rotation angle adjustment; And
Generating a composite image by synthesizing at least one final object mask obtained from the transformed at least one object mask on the selected background image; Including,
The transforming step
Obtaining at least one obstacle mask according to a predetermined obstacle probability value; And
Synthesizing the acquired obstacle mask with the transformed object mask so that the acquired obstacle mask overlaps on the transformed object mask to cover a part of the object mask to generate the final object mask; Including,
The transforming step
Adjusting the size and rotation angle of the obtained obstacle mask by applying a scaling parameter and a rotation parameter so that the obstacle mask does not overlay the object mask; Synthetic data generation method further comprising a.
The method of claim 7, wherein the transforming step
Selecting at least one object mask from among the plurality of object masks;
Selectively blurring the selected at least one object mask according to a predetermined blurring probability value; And
Outputting at least one deformed object mask by adjusting the size and rotation angle of the at least one object mask; Synthetic data generation method comprising a. The method of claim 8, wherein the blurring step
A method of generating composite data for blurring an object mask using a blur kernel that blurs the object mask according to a blur size and a blur direction angle. delete delete The method of claim 7, wherein generating the composite image comprises:
Placing at least one final object mask on the background image;
If the number of the final object masks is plural, adjusting the placement position so that the IOU (intersection of union) between the placed plural final object masks is less than a predetermined allowable value; And
Generating a verification label for each of at least one final object mask synthesized on the synthesized image; Synthetic data generation method comprising a.

Images