KR102268027B1 - Method for single image dehazing based on deep learning, recording medium and device for performing the method - Google Patents

Abstract

A deep learning-based image fog removal method includes: when a fog image is input, performing a convolution operation to extract image features; extracting n scaled image features by performing n (where n is a natural number) dilated convolution operations on the extracted image features at different rates; continuously performing a cumulative convolution operation in which the n scale image features are reflected as a cumulative input of the subsequent convolution operation through a plurality of residual dense blocks (RDBs); matching the n scaled image features on which the cumulative convolution operation is continuously performed with the same channel as the fog image; and outputting an image from which fog is removed by comparing and learning an image matched to the same channel as the fog image with the fog image. Accordingly, it is possible to obtain a fog removal image with improved quality even for fog images in various actual situations.

Description

A method for removing fog from a deep learning-based image, and a recording medium and apparatus for carrying out the same {METHOD FOR SINGLE IMAGE DEHAZING BASED ON DEEP LEARNING, RECORDING MEDIUM AND DEVICE FOR PERFORMING THE METHOD}

The present invention relates to a method for removing fog from a deep learning-based image, a recording medium and an apparatus for performing the same, and more particularly, to a computer vision technology for removing fog from an image obtained from a single camera.

Recently, with the development of imaging technology, various vision technologies are being developed. Among them, the fog removal technology removes the fog from the image by using the image obtained from the RGB-d camera and the images synthesized from the fog through the corresponding depth information as a database.

In the prior art, the conductivity, which is a variable constituting the model, is estimated based on the fog image model. In the first prior art (non-patent document 1 of the prior art document), conductivity was estimated using a deep learning structure composed of feature extraction, multi-scale mapping, local extrema, and nonlinear regression.

In the second prior art (non-patent document 2 of the prior art document), the conductivity was estimated using a deep learning structure composed of a coarse-scale network and a fine-scale network. Both techniques apply the estimated conductivity to the fog image model to obtain the fog-removed image.

However, conventional techniques are techniques for estimating conductivity based on a fog image model and then substituting the estimated conductivity into the model to remove fog. Accordingly, since the conventional techniques estimate the conductivity based on the fog image model, there is a limitation in that it is difficult to properly remove the fog for the actual fog image to which the corresponding model is difficult to apply.

In addition, the existing computer vision-related technology was developed on the assumption that the image acquired from the camera is an image captured in a clear weather environment. Due to this, the developed computer vision technology cannot be used properly for images shot in a foggy environment.

KR 10-2014-0142381 A KR 10-1445577 B1

Dehazenet, An end-to-end system for single image haze removal, 2016 IEEE Transactions on Image Processing Wenqi Ren et al., Single image dehazing via multi-scale convolutional neural networks, 2016 IEEE European conference on computer vision

Accordingly, it is an object of the present invention to provide a method for removing fog from a deep learning-based image that is not based on a fog image model.

Another object of the present invention is to provide a recording medium in which a computer program for performing the method for removing fog of the deep learning-based image is recorded.

Another object of the present invention is to provide an apparatus for performing the method of removing the fog of the deep learning-based image.

In accordance with an embodiment of the present invention for realizing the object of the present invention, there is provided a method for removing fog from a deep learning-based image, comprising: when a fog image is input, performing a convolution operation to extract image features; extracting n scaled image features by performing n (where n is a natural number) dilated convolution operations on the extracted image features at different rates; continuously performing a cumulative convolution operation in which the n scale image features are reflected as a cumulative input of the subsequent convolution operation through a plurality of residual dense blocks (RDBs); matching the n scaled image features on which the cumulative convolution operation is continuously performed with the same channel as the fog image; and outputting an image from which fog is removed by comparing and learning an image matched to the same channel as the fog image with the fog image.

In an embodiment of the present invention, in the extracting of the image feature by performing the convolution operation, the image feature may be extracted using a 3Х3 convolution operation layer.

In an embodiment of the present invention, in the extracting of the n scale image features, a convolution operation may be performed while skipping pixels corresponding to a corresponding rate.

In an embodiment of the present invention, the extracting of the n-scale image features may include extracting image features of four scales using four 3Х3 dilated convolution operation layers.

In an embodiment of the present invention, in the step of extracting the n scale image features, the four 3Х3 dilated convolution operation layers are characterized at a rate of 1, 2, 4, and 8, respectively. can be analyzed.

In an embodiment of the present invention, in the step of continuously performing the cumulative convolution operation, the cumulative convolution operation may be successively performed using three residual dense blocks (RDBs) for each of the n scale image features. .

In an embodiment of the present invention, the step of matching the n scale image features on which the cumulative convolution operation is continuously performed to the same channel as the fog image includes: n scale image features on which the cumulative convolution operation is continuously performed performing a concatenation operation; and performing two convolution operations to match the same channel as the fog image.

In a computer-readable storage medium according to an embodiment for realizing another object of the present invention, a computer program for performing the method for removing fog of the deep learning-based image is recorded.

In a deep learning-based image fog removal apparatus according to an embodiment for realizing another object of the present invention, when a fog image is input, a feature extraction unit extracts image features by performing a convolution operation ; A scale division unit that extracts n scaled image features by performing n (here, n is a natural number) dilated convolution operation on the image features extracted from the feature extraction unit at different rates ; Continuously performing a cumulative convolution operation in which the n scale image features extracted from the scale division unit reflect the data on which the convolution operation has been performed through a plurality of residual dense blocks (RDBs) as the cumulative input of the subsequent convolution. RDB continuation; a matching unit that matches the features of n scaled images on which the cumulative convolution operation is continuously performed to the same channel as the fog image; and a learning unit that compares and learns an image matched to the same channel as the fog image with the fog image, and outputs an image from which fog is removed.

In an embodiment of the present invention, the feature extraction unit may extract image features using a 3Х3 convolutional operation layer.

In an embodiment of the present invention, the scale division unit performs a convolution operation while skipping pixels as much as a corresponding rate, and uses four 3Х3 dilated convolution operation layers. It is possible to extract dog-scale image features.

In an embodiment of the present invention, the scale division unit may analyze the characteristics of the four 3Х3 dilated convolution operation layers at rates of 1, 2, 4, and 8, respectively.

In an embodiment of the present invention, the RDB continuation unit may successively perform a cumulative convolution operation on the n scale image features using three residual dense blocks (RDBs), respectively.

In an embodiment of the present invention, the matching unit may include: a concatenation operation unit configured to perform a concatenation operation on n scale image features on which the cumulative convolution operation is continuously performed; and a first and second convolution operation unit that performs two convolution operations to match the same channel as the fog image.

According to this deep learning-based image fog removal method, since it is not based on a fog image model, it is possible to obtain a fog removal image with improved quality even for fog images in various situations.

1 is a block diagram of an apparatus for removing fog of a deep learning-based image according to an embodiment of the present invention.
2 is a diagram for explaining a dilated convolution according to a rate in the present invention.
3 is a block diagram showing the structure of a Resisual Block used in deep learning of the present invention.
4 is a block diagram showing the structure of a Dense Block used in deep learning of the present invention.
5 is a block diagram showing the structure of a RDB (Resisual Dense Block) used in deep learning of the present invention.
6 is a flowchart of a method for removing fog from a deep learning-based image according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0023] Reference is made to the accompanying drawings, which show by way of illustration specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different but need not be mutually exclusive. For example, certain shapes, structures, and characteristics described herein with respect to one embodiment may be embodied in other embodiments without departing from the spirit and scope of the invention. In addition, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the present invention. Accordingly, the detailed description set forth below is not intended to be taken in a limiting sense, and the scope of the invention, if properly described, is limited only by the appended claims, along with all scope equivalents to those claimed. Like reference numerals in the drawings refer to the same or similar functions throughout the various aspects.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

1 is a block diagram of an apparatus for removing fog of a deep learning-based image according to an embodiment of the present invention. 2 is a diagram for explaining a dilated convolution according to a rate in the present invention.

The deep learning-based image fog removal device (10, hereinafter device) according to the present invention is a device for removing fog from an image acquired from a single camera, and is not based on a fog image model but can be applied to fog images in various situations Do.

Referring to FIG. 1 , an apparatus 10 according to the present invention includes a feature extraction unit 100 , scale division units 310 , 330 , 350 , 370 , and RDB continuation units 312 , 314 , 316 , 332 , 334 , 336 . , 352 , 354 , 356 , 372 , 374 , 376 , matching units 500 , 600 , 700 , and a learning unit 900 .

In the device 10 of the present invention, software (application) for performing fog removal of a deep learning-based image may be installed and executed, and the feature extraction unit 100, the scale division unit 310, 330, 350, 370), the RDB continuation unit 312, 314, 316, 332, 334, 336, 352, 354, 356, 372, 374, 376, the matching unit 500, 600, 700, and the learning unit 900 The configuration of may be controlled by software for performing fog removal of the deep learning-based image executed in the device 10 .

The device 10 may be a separate terminal or a module of the terminal. In addition, the feature extraction unit 100, the scale division unit 310, 330, 350, 370, and the RDB continuation unit 312, 314, 316, 332, 334, 336, 352, 354, 356, 372, 374 , 376), the matching units 500 , 600 , 700 , and the learning unit 900 may be configured as an integrated module, or may include one or more modules. However, on the contrary, each configuration may be formed of a separate module.

The device 10 may be movable or stationary. The apparatus 10 may be in the form of a server or an engine, and may be a device, an application, a terminal, a user equipment (UE), a mobile station (MS), or a wireless device. (wireless device), may be called other terms such as a handheld device (handheld device).

The device 10 may execute or manufacture various software based on an operating system (OS), that is, the system. The operating system is a system program for software to use the hardware of the device, and is a mobile computer operating system such as Android OS, iOS, Windows Mobile OS, Bada OS, Symbian OS, Blackberry OS, and Windows series, Linux series, Unix series, It can include all computer operating systems such as MAC, AIX, and HP-UX.

When a fog image is input, the feature extraction unit 100 extracts image features by performing a convolution operation. In an embodiment, image features may be extracted using a 3Х3 convolutional operation layer.

The scale division unit 310 , 330 , 350 , 370 divides the image features extracted from the feature extraction unit 100 into n (here, n is a natural number) dilated convolutions having different rates. convolution) operation to extract n scale image features.

In dilated convolution, the convolution operation is performed while skipping pixels as much as the corresponding rate. Since dilated convolution performs a convolution operation while skipping pixels by a rate, image features at various scales can be analyzed according to the rate.

Referring to FIG. 2 , when the rate is 1, convolution is performed at intervals of 1 pixel, and when the rate is 2, convolution is performed at intervals of 2 pixels. Similarly, when the rate is 3, convolution is performed at an interval of 3 pixels.

In an embodiment, each of the scale division units 310 , 330 , 350 , and 370 may extract image features of four scales using four 3Х3 dilated convolution operation layers.

In FIG. 1 , numbers in parentheses of the scale division units 310 , 330 , 350 , 370 indicate a rate of dilated convolution. However, the rate and the 3Х3 expanded convolution are only examples, and may be performed by changing the rate and variables of the nХn expanded convolution, if necessary.

The scale division unit 310 , 330 , 350 , 370 analyzes the characteristics of the four 3Х3 dilated convolution operation layers at rates of 1, 2, 4 and 8, respectively. Image features at scale can be analyzed.

Specifically, the first scale dividing unit 310 performs 3Х3 dilated convolution at a rate of 1, and the second scale dividing unit 330 expands 3Х3 at a rate of 2. Perform dilated convolution.

In addition, the third scale dividing unit 350 performs 3Х3 dilated convolution at a rate of 4, and the fourth scale dividing unit 370 performs 3Х3 dilated convolution at a rate of 8. Perform dilated convolution. Accordingly, in the present invention, image features at various scales can be analyzed.

The RDB continuation units 312, 314, 316, 332, 334, 336, 352, 354, 356, 372, 374, and 376 are n scale images extracted from the scale division units 310, 330, 350, and 370. The cumulative convolution operation, which reflects the data on which the convolution operation has been performed through a plurality of residual dense blocks (RDBs) for each feature, as a cumulative input of the subsequent convolution, is continuously performed.

3 is a block diagram showing the structure of a Resisual Block used in deep learning of the present invention, and FIG. 4 is a block diagram showing the structure of a Dense Block used in deep learning of the present invention. 5 is a block diagram showing the structure of a RDB (Resisual Dense Block) used in deep learning of the present invention.

The present invention uses the residual dense block (RDB) of FIG. 5 as a basic structure. The structure of the RDB is a structure in which the residual block of FIG. 3 and the dense block of FIG. 4 are combined through convolution and ReLU. RDB is a proposed structure to solve the super-resolution problem of images.

Here, the cumulative convolution refers to performing cumulative convolution that is reflected as a cumulative input of subsequent convolution, and as shown in FIG. 4 , the output of the first RDB synthesis block is a second RDB synthesis block, a third RDB synthesis block, and four It becomes the input of the first RDB synthesis block. Similarly, the output of the second RDB composition block becomes the input of the third RDB composition block and the fourth RDB composition block, and the output of the third RDB composition block becomes the input of the fourth RDB composition block.

In the present invention, it is possible to further analyze the features on the corresponding scale by using a residual dense block (RDB) for each of the features obtained through the four types of expanded convolution. RDB is a module constructed using residual blocks and dense blocks, and is very effective for feature analysis.

In an embodiment of the present invention, a total of 12 RDBs were used by using 3 RDBs for each scale. Specifically, the output of the first scale division unit 310 passes through the RDB continuation units 312 , 314 , and 316 and sequentially performs the cumulative convolution operation using three residual dense blocks (RDBs).

The output of the second scale division unit 330 goes through the RDB continuation units 332 , 334 , and 336 to successively perform a cumulative convolution operation using three RDBs (residual dense blocks), and the third scale division unit ( The output of 350 goes through the RDB continuation units 352 , 354 , and 356 , and a cumulative convolution operation is continuously performed using three residual dense blocks (RDBs). Similarly, the output of the fourth scale division unit 370 passes through the RDB continuation units 372 , 374 , and 316 and sequentially performs a cumulative convolution operation using three residual dense blocks (RDBs).

The matching units 500 , 600 , and 700 match the features of the n scale images that have been successively performed the cumulative convolution operation to the same channel as the fog image. Specifically, the matching unit 500 , 600 , 700 includes a concatenation operation unit 500 that performs a concatenation operation on n scale image features on which the cumulative convolution operation is continuously performed and the same channel as the fog image. It may include first and second convolution operation units 600 and 700 that perform two convolution operations to match with .

The learning unit 900 compares and learns an image matched to the same channel as the fog image with the fog image, and outputs an image from which fog is removed.

In order to carry out the present invention, 3733 images of clear weather and corresponding depth images were collected, and each image was synthesized in 5 ways to create a total of 18665 images of foggy weather and used as a learning database. did. In addition, for a part of the same sunny weather image and corresponding depth images, 1449 images were collected and used to verify the performance of the learned network by collecting fog-synthesized images by other methods.

In the present invention, to verify the performance of the network, fog is removed from 1449 fog images, and PSNR and SSIM between the fog-removed image and the clear image are compared. As a result of the performance verification, PSNR is 21.11 and SSIM is 0.88, which shows high performance compared to prior art 1 (PSNR: 12.84, SSIM: 0.72) and prior art 2 (PSNR: 12.27, SSIM: 0.7).

Since the conventional technology estimates the conductivity based on the fog image model, there is a limit to properly removing the fog for the actual fog image, which is difficult to apply the model, but the technology developed in the present invention is not based on the fog image model. It can bring better results even for fog images in various actual situations.

In addition, the existing computer vision-related technology was developed on the assumption that the image acquired from the camera is an image captured in a clear weather environment. Due to this, the developed computer vision technology cannot be used properly for images shot in a foggy environment. Accordingly, improved performance may be obtained by applying the previously developed computer vision technology after applying the fog removal technology developed in the present invention.

6 is a flowchart of a method for removing fog from a deep learning-based image according to an embodiment of the present invention.

The deep learning-based image mist removal method according to the present embodiment may be performed in substantially the same configuration as the apparatus 10 of FIG. 1 . Accordingly, the same components as those of the device 10 of FIG. 1 are given the same reference numerals, and repeated descriptions are omitted.

In addition, the deep learning-based image fog removal method according to the present embodiment may be executed by software (application) for performing the deep learning-based image fogging.

The deep learning-based image fog removal method according to the present invention is a technique for removing fog from an image acquired from a single camera, and is not based on a fog image model but can be applied to fog images in various situations.

Referring to FIG. 6 , in the deep learning-based image fog removal method according to the present embodiment, when a fog image is input, a convolution operation is performed to extract image features (step S10). In an embodiment, image features may be extracted using a 3Х3 convolutional operation layer.

n (here, n is a natural number) dilated convolution operations having different rates on the extracted image features to extract n scaled image features (step S20).

In dilated convolution, the convolution operation is performed while skipping pixels as much as the corresponding rate. Since dilated convolution performs a convolution operation while skipping pixels by a rate, image features at various scales can be analyzed according to the rate.

For example, when the rate is 1, convolution is performed at intervals of 1 pixel, and when the rate is 2, convolution is performed at intervals of 2 pixels. Similarly, when the rate is 3, convolution is performed at intervals of 3 pixels (see FIG. 2 ).

In an embodiment, four scale image features may be extracted using four 3Х3 dilated convolution operation layers. However, the rate and the 3Х3 expanded convolution are only examples, and may be performed by changing the rate and variables of the nХn expanded convolution, if necessary.

The four 3Х3 dilated convolution operation layers may analyze features at a rate of 1, 2, 4, and 8, respectively, to analyze image features at four scales. Accordingly, in the present invention, image features at various scales can be analyzed.

A cumulative convolution operation is continuously performed in which the data on which the convolution operation has been performed on the n scale image features through a plurality of residual dense blocks (RDBs) is reflected as a cumulative input of the subsequent convolution (step S30).

A cumulative convolution operation is continuously performed in which the extracted n scale image features are reflected as the cumulative input of the convolution operation through a plurality of residual dense blocks (RDBs).

The present invention uses the residual dense block (RDB) of FIG. 5 as a basic structure. The structure of the RDB is a structure in which the residual block of FIG. 3 and the dense block of FIG. 4 are combined through convolution and ReLU. RDB is a proposed structure to solve the super-resolution problem of images.

Here, the cumulative convolution refers to performing cumulative convolution that is reflected as a cumulative input of subsequent convolution, and as shown in FIG. 4 , the output of the first RDB synthesis block is a second RDB synthesis block, a third RDB synthesis block, and four It becomes the input of the first RDB synthesis block. Similarly, the output of the second RDB composition block becomes the input of the third RDB composition block and the fourth RDB composition block, and the output of the third RDB composition block becomes the input of the fourth RDB composition block.

In the present invention, it is possible to further analyze the features on the corresponding scale by using a residual dense block (RDB) for each of the features obtained through the four types of expanded convolution. RDB is a module constructed using residual blocks and dense blocks, and is very effective for feature analysis.

In an embodiment of the present invention, a total of 12 RDBs were used by using 3 RDBs for each scale.

The n scale image features on which the cumulative convolution operation is continuously performed are matched with the same channel as the fog image (step S40). Step S40 is a step of performing a concatenation operation on n scale image features that have been successively performed the cumulative convolution operation, and performing two convolution operations to match the same channel as the fog image. may include steps.

An image matched with the same channel as the fog image is compared and learned with the fog image (step S50), and an image from which fog is removed is output.

Accordingly, since the technology developed in the present invention is not based on a fog image model, it can bring better results for fog images in various situations.

Such a deep learning-based image fog removal method may be implemented as an application or implemented in the form of program instructions that may be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc. alone or in combination.

The program instructions recorded on the computer-readable recording medium are specially designed and configured for the present invention, and may be known and available to those skilled in the computer software field.

Examples of the computer-readable recording medium include a hard disk, a magnetic medium such as a floppy disk and a magnetic tape, an optical recording medium such as a CD-ROM, a DVD, and a magneto-optical medium such as a floppy disk. media), and hardware devices specially configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.

Although described above with reference to the embodiments, those skilled in the art can variously modify and change the present invention within the scope without departing from the spirit and scope of the present invention described in the claims below You will understand.

The present invention is a computer vision technology that removes the fog of an image acquired from a single camera, and can be usefully applied to a smartphone application and the like. In addition, it can be applied to various fields such as automobile industry, agricultural industry, and surveillance camera.

Since the conventional technology estimates the conductivity based on the fog image model, there is a limit to properly removing the fog for the actual fog image, which is difficult to apply the model, but the technology developed in the present invention is not based on the fog image model. It can bring better results even for fog images in various actual situations.

In addition, the existing computer vision-related technology was developed on the assumption that the image acquired from the camera is an image captured in a clear weather environment. Due to this, the developed computer vision technology cannot be used properly for images shot in a foggy environment. Therefore, improved performance can be obtained by applying the previously developed computer vision technology after applying the fog removal technology developed in this development.

10: Deep learning-based video fog removal device
100: feature extraction unit
310, 330, 350, 370: scale division
312, 314, 316: RDB continuation
500: joint operation unit
600: first convolution operation unit
700: second convolution operation unit
900: study department

Claims (14)

when a fog image is input, performing a convolution operation to extract image features;
extracting n scaled image features by performing n (where n is a natural number) dilated convolution operations on the extracted image features at different rates;
continuously performing a cumulative convolution operation in which the data on which the convolution operation has been performed through a plurality of residual dense blocks (RDBs) for each of the n scale image features is reflected as a cumulative input of the subsequent convolution;
matching n scaled image features on which the cumulative convolution operation is continuously performed with the same channel as the fog image; and
Comparing and learning the image matched to the same channel as the fog image with the fog image, and outputting an image from which fog is removed;
The step of extracting the n scale image features,
Extracting four scale image features using four 3Х3 dilated convolution operation layers,
In order to analyze image features at various scales according to different ratios, the four 3Х3 dilated convolution operation layers generate pixels as many as 1, 2, 4, and 8 rates, respectively. It uses a convolution operation performed by skipping,
The step of continuously performing the cumulative convolution operation comprises:
A method for removing fog in a deep learning-based image, in which each of the extracted four scale image features is reflected as a cumulative input of a convolution using three residual dense blocks (RDBs).
The method of claim 1, wherein the extracting of image features by performing the convolution operation comprises:
A deep learning-based image fog removal method that extracts image features using a 3Х3 convolutional operation layer.
delete delete delete delete The method of claim 1, wherein the matching of n scale image features on which the cumulative convolution operation is continuously performed to the same channel as the fog image comprises:
performing a concatenation operation on n scale image features on which the cumulative convolution operation is continuously performed; and
Performing two convolution operations to match the same channel as the fog image; Containing, Deep learning-based image fog removal method.
A computer-readable storage medium in which a computer program for performing the method for removing fog of the deep learning-based image according to any one of claims 1 and 7 is recorded.
a feature extracting unit that extracts image features by performing a convolution operation when a fog image is input;
A scale dividing unit extracting n scaled image features by performing n (here, n is a natural number) dilated convolution operations on the image features extracted from the feature extractor having different rates. ;
Continuously performing a cumulative convolution operation in which the n scale image features extracted from the scale division unit reflect the data on which the convolution operation has been performed through a plurality of residual dense blocks (RDBs) as the cumulative input of the subsequent convolution. RDB continuation;
a matching unit that matches the features of n scaled images on which the cumulative convolution operation is continuously performed to the same channel as the fog image; and
A learning unit that compares and learns an image matched to the same channel as the fog image with the fog image, and outputs an image from which the fog is removed;
The scale division unit,
Extracting four scale image features using four 3Х3 dilated convolution operation layers,
In order to analyze image features at various scales according to different ratios, the four 3Х3 dilated convolution operation layers each have a corresponding rate of 1, 2, 4, and 8 pixels. It uses the convolution operation performed while skipping
The RDB continuation part,
An apparatus for removing fog of a deep learning-based image, which is reflected as a cumulative input of a convolution using three residual dense blocks (RDB) for each of the extracted four scale image features.
The method of claim 9, wherein the feature extraction unit,
A device that extracts image features using 3Х3 convolutional operation layer and removes fog from deep learning-based images.
delete delete delete The method of claim 9, wherein the matching unit,
a concatenation operation unit configured to perform a concatenation operation on n scale image features that have been successively performed the cumulative convolution operation; and
First and second convolution operation units that perform two convolution operations to match the same channel as the fog image; and a deep learning-based image fog removal apparatus.

Images