KR102251799B1 - Non-supervised-learning-based poor vehicle number image recovery learning method and device - Google Patents

Abstract

This non-teacher learning-based poor vehicle number image recovery learning device is a low-quality noise control learning device that receives low-quality image data with poor identification of the license plate number, outputs low-quality noise-removing image data from which noise is removed, and then determines a learning loss. part; When high-definition image data capable of identifying the license plate number and the low-quality noise-removing image data are input, the low-quality noise-removing image data is matched with the low-quality recovered image data to output the converted high-definition image data, and then learning loss for matching ( A high-definition image learning unit to determine Loss) And an image determination unit configured to determine a training loss after determining whether the low-quality recovered image data provided through the high-quality image learning unit is high-quality image data or low-quality recovered image data.

Description

{Non-supervised-learning-based poor vehicle number image recovery learning method and device}

The present invention relates to a learning method and apparatus for recovering a poor vehicle number image based on non-teacher learning, and more particularly, a learning method for improving and recovering a low-quality vehicle number image so that vehicle number can be identified from an image photographed by a camera. And a device.

In general, the most required information to track the vehicle of the dragon is the vehicle number. These vehicle numbers can be obtained through CCTV installed on the road.

However, thanks to the advancement of technology, if a high-definition CCTV is installed, it can automatically recognize and identify the license plate of a vehicle, but most roads do not have high-definition CCTV yet, and low-quality general security CCTV or near the accident site. In many cases, the black box video taken from the vehicle is all.

It is difficult to identify the car number for the vehicle in question because of low-quality general CCTVs or black boxes installed in vehicles are of poor quality. In particular, images taken at night, images taken in bad weather, or images with severe motion blur are often difficult to identify even when a person sees them with the naked eye.

The present invention was conceived to solve the conventional problem, and to provide a non-teacher learning-based poor vehicle number image recovery learning method and apparatus for recovering a vehicle number image photographed in a poor environment into a readable image using deep learning technology. do.

The object of the present invention is not limited to the above-mentioned object, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

In order to achieve the above object, the apparatus for learning to recover poor vehicle number image based on non-teacher learning according to an embodiment of the present invention receives low-quality image data in which the number identification of the vehicle license plate is low, and outputs low-quality noise-removing image data from which noise is removed. A low-quality noise control learning unit that determines a post-learning loss; When high-definition image data capable of identifying the license plate number and the low-quality noise-removing image data are input, the low-quality noise-removing image data is matched with the low-quality recovered image data to output the converted high-definition image data, and then learning loss for matching ( A high-definition image learning unit to determine Loss) And an image determination unit configured to determine a training loss after determining whether the low-quality recovered image data provided through the high-quality image learning unit is high-quality image data or low-quality recovered image data.

It is preferable that the low-quality noise reduction model unit uses an auto-encoder that detects feature points while converting multidimensional input data into a low-dimensional code, and converting the low-dimensional code into first input multidimensional data.

In the method of recovering poor vehicle number image based on non-teacher learning according to an embodiment of the present invention, after the low-quality noise control learning unit receives low-quality image data in which the number identification of the license plate is low, and outputs low-quality noise-removing image data from which noise has been removed. Determining a learning loss; When the high-definition image learning unit outputs the converted high-definition image data by matching the low-quality noise-removing image data and the low-quality noise-removing image data for identification of the number of the license plate and the low-quality noise-removing image data. Determining a loss for learning about; And determining, by the image determination unit, whether the low-quality recovered image data provided through the high-quality image learning unit is high-quality image data or low-quality recovered image data, and then determining a training loss.

According to another embodiment of the present invention, an apparatus for recovering a poor vehicle number image based on non-teacher learning includes: a low-quality noise control learning unit configured to receive low-quality image data in which the number identification of the license plate is low and output low-quality noise-removing image data from which noise is removed; And a high-definition image learning unit that, when the low-quality noise-removing image data is input, matches the low-quality noise-removed image data with the low-quality restored image data, and outputs the converted high-quality image data.

According to an embodiment of the present invention, by recovering a vehicle number image photographed in a poor environment, the image quality of a vehicle number image that is difficult to read can be improved with a deep learning technique, and an effect of enabling a suitable vehicle tracking in a short time. There is.

As described above, according to an embodiment of the present invention, after low-quality image data that is difficult to identify the vehicle number is input, in the low-quality recovery data processed for image improvement by a deep learning method, it is possible to identify the vehicle number.

1 is a block diagram showing a configuration block of a learning apparatus for recovering an image of a poor vehicle number based on non-teacher learning according to an embodiment of the present invention.
2 is a reference diagram for explaining a learning result through a learning device for recovering poor vehicle number image based on non-teacher learning according to an embodiment of the present invention.
3 is a flowchart illustrating a method of learning to recover an image of a poor vehicle number based on non-teacher learning according to an embodiment of the present invention.
FIG. 4 is a block diagram illustrating an apparatus for recovering an image of a poor vehicle number based on non-teacher learning according to another embodiment of the present invention.

Advantages and features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various forms different from each other, and only these embodiments make the disclosure of the present invention complete, and common knowledge in the technical field to which the present invention pertains. It is provided to completely inform the scope of the invention to the possessor, and the invention is only defined by the scope of the claims. Meanwhile, terms used in the present specification are for explaining embodiments and are not intended to limit the present invention. In this specification, the singular form also includes the plural form unless specifically stated in the phrase. As used in the specification, "comprises" and/or "comprising" refers to the presence of one or more other components, steps, operations and/or elements in which the recited component, step, operation and/or element is Or does not preclude additions.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 is a block diagram illustrating an apparatus for learning to recover an image of a poor vehicle number based on non-teacher learning according to an embodiment of the present invention.

As shown in Figure 1, the non-teacher learning-based poor vehicle number image recovery learning apparatus according to an embodiment of the present invention is a low-quality noise control learning unit 100, a high-definition image learning unit 200, and an image determination unit 300 ).

The low-quality noise control learning unit 100 receives low-quality image data in which the number identification of the license plate is low, outputs low-quality noise-removing image data from which noise is removed, and then determines a training loss. On the other hand, it is preferable that the low-quality noise reduction model unit uses an auto encoder that detects feature points while converting multidimensional input data into a low-dimensional code and converting the low-dimensional code into the first input multidimensional data.

Meanwhile, the loss of the model determined by the low-quality noise control learning unit 100 can be expressed as shown in [Equation 1] below.

Here, "LQ_image" is low-quality image data, "~LQ_image" is low-quality noise removal image data, and "batch_size" represents the number of samples input to the model at one time.

The high-definition image learning unit 200 outputs the converted high-definition image data by matching the low-quality noise-removing image data to the low-quality restored image data when high-quality image data capable of identifying the number of the license plate and the low-quality noise-removing image data are input. Then, the learning loss for matching is determined.

Meanwhile, the loss of the model determined by the high-definition image learning unit 200 may be expressed as [Equation 2] below.

Here, "~HQ_image" is high-quality processed image data processed through the high-definition image learning unit 200 and then output, "HQ_image" is high-quality image data input to the high-quality image learning unit 200, and "~~LQ_image" "Is the low-quality recovered image data processed and output through the high-definition image learning unit 200, "~LQ_image" is the low-quality noise removal image data, and "HQ_valid" is processed and output through the image determination unit 300 High-quality image data.

In addition, the image determination unit 300 determines whether the low-quality recovered image data provided through the high-quality image learning unit 200 is high-quality image data or low-quality recovered image data, and then determines a training loss.

Meanwhile, the loss of the model determined by the image determination unit 300 may be expressed as [Equation 3] below.

Here, "~HQ_image" is high-quality image data processed and output through the high-definition image learning unit 200, and "HQ_valid" is high-quality image determination data processed and output through the image determination unit 300, and " ~~LQ_image" is the low-quality recovered image data processed by the high-definition image learning unit 200 and then output, and "HQ_fake" is the low-quality recovered image determination data processed and output by the image determination unit 300.

As described above, according to an embodiment of the present invention, by recovering a vehicle number image photographed in a poor environment, the image quality of a vehicle number image that is difficult to read can be improved with a deep learning technique, and a vehicle tracking of the vehicle is possible in a short time. It has the effect of making it happen.

That is, according to an embodiment of the present invention, as shown in FIG. 2, it can be confirmed that improved image data can be obtained.

The first image data (L) is the low-quality image data, the next image data (LD) is the low-quality noise-removing image data, the next image data (LG) is the low-quality recovery data, and the last image data (H) is the original image data. This is the number of correct answers.

In this way, after low-quality image data that is difficult to identify the vehicle number is input, in the low-quality recovery data processed for image improvement by a deep learning method, it is possible to identify the vehicle number.

3 is a flowchart illustrating a method of learning to recover a poor vehicle number image based on non-teacher learning according to an embodiment of the present invention.

Such, the non-teacher learning-based poor vehicle number image recovery learning method according to an embodiment of the present invention is a learning device including a low-quality noise control learning unit 100, a high-definition image learning unit 200, and an image determination unit 300 It is preferably carried out by.

The low-quality noise control learning unit 100 receives low-quality image data in which the number identification of the license plate is low, outputs low-quality noise-removing image data from which noise is removed, and then determines a training loss (S100). Here, it is preferable that the low-quality noise reduction model unit uses an auto encoder that detects feature points while converting the multidimensional input data into a low-dimensional code and converting the low-dimensional code into the first input multidimensional data.

Meanwhile, the loss of the model determined by the low-quality noise control learning unit 100 may be expressed as in [Equation 1].

Subsequently, when high-quality image data capable of identifying the number of the license plate and the low-quality noise-removing image data are input, the high-quality image learning unit 200 matches the low-quality noise-removing image data to the low-quality restored image data and converted high-quality image data. After outputting is determined, a learning loss for matching is determined (S200).

Meanwhile, the loss of the model determined by the high-definition image learning unit 200 may be expressed as in [Equation 2].

Thereafter, the image determination unit 300 determines whether the low-quality recovered image data provided through the high-quality image learning unit 200 is high-quality image data or low-quality recovered image data, and then determines a training loss (S300).

Meanwhile, the loss of the model determined by the image determination unit 300 may be expressed as in [Equation 3].

According to an embodiment of the present invention, by recovering a vehicle number image photographed in a poor environment, the image quality of a vehicle number image that is difficult to read can be improved with a deep learning technique, and an effect of enabling a suitable vehicle tracking in a short time. There is.

That is, according to an embodiment of the present invention, as shown in FIG. 2, it can be confirmed that improved image data can be obtained.

The first image data (L) is the low-quality image data, the next image data (LD) is the low-quality noise-removing image data, the next image data (LG) is the low-quality recovery data, and the last image data (H) is the original image data. This is the number of correct answers.

In this way, after low-quality image data that is difficult to identify the vehicle number is input, in the low-quality recovery data processed for image improvement by a deep learning method, it is possible to identify the vehicle number.

FIG. 4 is a block diagram illustrating an apparatus for recovering an image of a poor vehicle number based on non-teacher learning according to another embodiment of the present invention.

As shown in FIG. 4, an apparatus for recovering a poor vehicle number image based on non-teacher learning according to another embodiment of the present invention includes a low-quality noise control processing unit 400 and a high-quality image processing unit 500.

The low-quality noise control processing unit 400 receives low-quality image data in which the number identification of the license plate is low by using the deep learning technique learned through the low-quality noise control learning unit 100, and outputs the low-quality noise-removing image data from which noise is removed. do.

When the low-quality noise-removing image data is input using the deep learning technique learned through the high-definition image learning unit 200, the high-quality image processing unit 500 matches the low-quality noise-removing image data to the low-quality recovered image data, and is converted. Output high-quality image data.

That is, according to an embodiment of the present invention, as shown in FIG. 2, it can be confirmed that improved image data can be obtained.

The first image data (L) is the low-quality image data, the next image data (LD) is the low-quality noise-removing image data, the next image data (LG) is the low-quality recovery data, and the last image data (H) is the original image data. This is the number of correct answers.

In this way, after low-quality image data that is difficult to identify the vehicle number is input, in the low-quality recovery data processed for image improvement by a deep learning method, it is possible to identify the vehicle number.

delete

In the above, the configuration of the present invention has been described in detail with reference to the accompanying drawings, but this is only an example, and if one of ordinary skill in the art to which the present invention belongs, various modifications and changes within the scope of the technical idea of the present invention Of course this is possible. Therefore, the scope of protection of the present invention should not be limited to the above-described embodiments, but should be determined by the description of the following claims.

100: low-quality noise control learning unit 200: high-quality image learning unit
300: image judgment unit

Claims (12)

A low-quality noise control learning unit that receives low-quality image data from which the number identification of the license plate is low, outputs low-quality noise-removing image data from which noise is removed, and then determines a training loss using the low-quality image data and low-quality noise-removing image data;
When high-quality image data capable of identifying the license plate number and the low-quality noise-removing image data are input, the converted high-quality image data is output by matching the low-quality noise-removing image data to the low-quality restored image data, and then processed and output high-quality processed image. A high-definition image learning unit that determines a training loss using data, high-quality image data, low-quality recovered image data, and low-quality noise-removed image data; And
High-definition image data output after determining whether the low-quality recovered image data provided through the high-definition image learning unit is high-definition image data or low-quality recovered image data, and output low-quality recovered image data processed through the high-quality image learning unit And an image determination unit for determining a learning loss using the low-quality recovered image determination data.
The method of claim 1,
The low-quality noise control learning unit,
Non-teacher learning-based poor vehicle number image recovery learning apparatus using an auto encoder that detects feature points while changing multidimensional input data into low-dimensional codes and converting the low-dimensional codes back to first input multidimensional data.
The method of claim 1,
The loss of the model determined by the low-quality noise control learning unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning apparatus.
[Equation 1]

Here, "LQ_image" is low-quality image data, "~LQ_image" is low-quality noise removal image data, and "batch_size" is the number of samples input to the model at one time.
The method of claim 1,
The loss of the model determined by the high-definition image learning unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning device.
[Equation 2]

Here, "~HQ_image" is high-definition processed image data processed and output through the high-definition image learning unit, "HQ_image" is high-quality image data input to the high-definition image learning unit, and "~~LQ_image" is through the high-definition image learning unit. This is the low-quality recovered image data output after processing, "~LQ_image" is the low-quality noise removal image data, and "HQ_valid" is the high-quality image data processed and output by the image determination unit.
The method of claim 1,
The loss of the model determined by the image determination unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning device.

Here, "~HQ_image" is high-quality image data processed and output by the high-definition image learning unit, "HQ_valid" is high-quality image judgment data processed and output by the image determination unit, and "~~LQ_image" is high-quality image learning. This is the low-quality recovered image data processed by the unit and then output, and "HQ_fake" is the low-quality recovered image determination data processed and output through the image determination unit.
The method of claim 1,
A low-definition noise control processor configured to receive low-quality image data in which the number identification of the license plate is low by using the deep learning technique learned through the low-quality noise control learning unit and output low-quality noise-removing image data from which noise is removed; And
When the low-definition noise-removing image data is input using the deep learning technique learned through the high-definition image learning unit, the high-definition image processing unit further outputs the converted high-definition image data by matching the low-quality noise-removing image data to the low-quality restored image data. Poor vehicle number image recovery learning device based on non-teacher learning including.
Receiving, by a low-quality noise control learning unit, low-quality image data in which the number identification of the license plate is lowered, outputting low-quality noise-removing image data from which noise is removed, and then determining a training loss using the low-quality image data and the low-quality noise-removing image data;
After the high-definition image learning unit outputs the converted high-definition image data by matching the low-quality noise-removing image data and the low-quality recovered image data when the high-definition image data and the low-quality noise-removing image data that can identify the license plate number are input. Determining a training loss using the output high-quality processed image data, high-quality image data, low-quality restored image data, and low-quality noise-removed image data; And
The image determination unit determines whether the low-quality recovered image data provided through the high-definition image learning unit is high-quality image data or low-quality recovered image data, and then output high-quality image data, high-quality image determination data, and low-quality output after being processed through the high-quality image learning unit. Determining a training loss using the recovered image data and the low-quality recovered image determination data; Non-teacher learning-based poor vehicle number image recovery learning method comprising a.
The method of claim 7,
The low-quality noise control learning step,
Non-teacher learning-based poor vehicle number image recovery learning method using an auto encoder that detects feature points while changing multidimensional input data into low-dimensional codes and converting the low-dimensional codes back to first input multidimensional data.
The method of claim 7,
The loss of the model determined by the low-quality noise control learning unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning method.
[Equation 1]

Here, "LQ_image" is low-quality image data, "~LQ_image" is low-quality noise removal image data, and "batch_size" is the number of samples input to the model at one time.
The method of claim 7,
The loss of the model determined by the high-definition image learning unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning method.
[Equation 2]

Here, "~HQ_image" is high-definition processed image data processed and output through the high-definition image learning unit, "HQ_image" is high-quality image data input to the high-definition image learning unit, and "~~LQ_image" is through the high-definition image learning unit. This is the low-quality recovered image data output after processing, "~LQ_image" is the low-quality noise removal image data, and "HQ_valid" is the high-quality image data processed and output by the image determination unit.
The method of claim 7,
The loss of the model determined by the image determination unit is represented by the following equation. Non-teacher learning-based poor vehicle number image recovery learning method.

Here, "~HQ_image" is high-quality image data processed and output by the high-definition image learning unit, "HQ_valid" is high-quality image judgment data processed and output by the image determination unit, and "~~LQ_image" is high-quality image learning. This is the low-quality recovered image data processed by the unit and then output, and "HQ_fake" is the low-quality recovered image determination data processed and output through the image determination unit.
The method of claim 7,
A low-quality noise control processing step of receiving low-quality image data in which the number identification of the license plate is low by using the deep learning technique learned through the low-quality noise control learning unit and outputting low-quality noise-removing image data from which noise has been removed; And
Non-teacher learning base further comprising a high-definition image processing step of outputting the converted high-definition image data by matching the low-quality noise-removing image data input using the deep learning technique learned through the high-definition image learning unit with low-quality recovered image data. How to learn to recover poor vehicle number image.

Images