KR20220125124A - Image processing apparatus and operating method for the same - Google Patents

Abstract

The present invention relates to an image processing apparatus to perform adaptive image processing according to the characteristics of each pixel included in an image. According to the present invention, the image processing apparatus comprises: a memory storing one or more instructions; and a processor executing the one or more instructions stored in the memory. The processor obtains similarity information indicating a similarity between each of pixels included in a first image and neighboring pixels of each of the pixels; generates a weight map including weight information corresponding to each pixel on the basis of the similarity information; generates a spatially variant kernel including a kernel corresponding to each pixel on the basis of a spatial kernel including weight information according to a positional relationship between each pixel and a neighboring pixel, and a weight map; and applies the spatially variant kernel to the first image to generate a second image.

Description

Image processing apparatus and operating method for the same

Various embodiments relate to an image processing apparatus for improving image quality by using a neural network, and an operating method thereof.

As data traffic increases exponentially with the development of computer technology, artificial intelligence has become an important trend driving future innovation. Since artificial intelligence is a method that imitates the way of thinking of humans, it can be applied infinitely to virtually all industries. Representative technologies of artificial intelligence include pattern recognition, machine learning, expert systems, neural networks, and natural language processing.

A neural network models the characteristics of human biological nerve cells by mathematical expressions, and uses an algorithm that mimics the ability of learning that humans have. Through this algorithm, the neural network can generate a mapping between input data and output data, and the ability to generate this mapping can be expressed as the learning ability of the neural network. In addition, the neural network has a generalization ability to generate correct output data with respect to input data that has not been used for learning, based on the learned result.

When image processing such as denoising of an image is performed using a deep neural network (eg, a deep convolutional neural network (CNN)), each pixel included in the image is When the same kernel (filter) is applied, there is a problem in that image processing performance is degraded. Therefore, during image processing, it is necessary to perform image processing by applying different kernels according to positions or intensity characteristics of pixels included in an image.

Various embodiments may provide an image processing apparatus capable of performing adaptive image processing according to characteristics of each pixel included in an image using a convolutional neural network, and an operating method thereof.

An image processing apparatus according to an embodiment includes a memory storing one or more instructions, and a processor executing the one or more instructions stored in the memory, wherein the processor comprises: each of pixels included in a first image and the Obtaining similarity information indicating the similarity of each of the pixels with the neighboring pixel, generating a weight map including weight information corresponding to each of the pixels based on the similarity information, and generating each of the pixels and the neighboring pixel generating a spatially variable kernel including a kernel corresponding to each of the pixels based on a spatial kernel including weight information according to a positional relationship with By applying a spatially variant kernel, the second image may be generated.

The processor according to an embodiment obtains first similarity information based on a difference between each of the pixels included in the first image and a first neighboring pixel having a first relative position with respect to each of the pixels, , second similarity information may be obtained based on a difference between each of the pixels included in the first image and a second neighboring pixel having a second relative position with respect to each of the pixels.

The processor according to an embodiment may generate the weight map by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using the convolutional neural network.

The number of channels of at least one of the similarity information, the weight map, and the spatially variable kernel according to an embodiment may be determined based on the size of the spatial kernel.

In the spatial kernel according to an embodiment, a pixel located at the center of the spatial kernel may have a largest value, and a pixel value may decrease as it moves away from the center.

The size of the spatial kernel according to an embodiment is K x K, the number of channels of the weight map is K ² , and the processor arranges pixel values included in the spatial kernel in the channel direction to 1 x 1 x K The spatially variable kernel can be generated by converting the weight vector into a weight vector having a size of ² and performing a multiplication operation of the weight vector with each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the weight map. can

The spatially variable kernel according to an embodiment may include the same number of kernels as the number of pixels included in the first image.

The processor according to an embodiment performs filtering by applying a first kernel included in the spatially variable kernel to a first region centered on a first pixel included in the first image, and performs filtering on the first image. The second image may be generated by applying a second kernel included in the spatial transformation kernel to a second region centered on the included second pixel and performing filtering.

The processor according to an embodiment generates a first weight map by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using a convolutional neural network, and The weight map may be generated by performing a dilated convolution operation between the first weight map and the first kernel.

According to an embodiment, the number of channels in the first weight map and the number of channels in the first kernel are the same, and the processor performs a convolution operation in which the channels corresponding to the first weight map and the first kernel are expanded ( By performing depthwise dilated convolution), the weighted map can be generated.

The method of operating an image processing apparatus according to an embodiment may include: obtaining similarity information indicating a similarity between each pixel included in a first image and a neighboring pixel of each of the pixels; based on the similarity information, the generating a weight map including weight information corresponding to each of the pixels; based on the weight map and a spatial kernel including weight information according to a positional relationship between each of the pixels and the neighboring pixels , generating a spatially variable kernel including a kernel corresponding to each of the pixels, and generating a second image by applying the spatially variant kernel to the first image. have.

In an image processing apparatus according to an embodiment, when image processing is performed on each of a plurality of pixels included in an image by using similarity information between each pixel and neighboring pixels, the image processing is performed on neighboring pixels having similar characteristics. characteristics can be further emphasized.

The image processing apparatus according to an embodiment may apply different kernels generated based on the regional characteristics of the pixel to each of the pixels included in the image, thereby enabling adaptive image processing according to the regional characteristics of the pixel. .

The image processing apparatus according to an embodiment may perform detailed edge processing of an input image and denoising, which removes noise while maintaining a texture, using a convolutional neural network.

1 is a diagram illustrating an operation of an image processing apparatus processing an image using an image processing network according to an exemplary embodiment.
2 is a diagram illustrating an image processing network according to an exemplary embodiment.
3 is a diagram referenced to describe a method of generating similarity information according to an embodiment.
4 and 5 are diagrams referenced to describe methods of obtaining similarity information according to an embodiment.
6 is a diagram referenced to describe a method of generating a weighted map according to an embodiment.
7 and 8 are diagrams referenced to describe a convolution operation performed in a convolution layer according to an embodiment.
9 is a diagram referenced to describe a method of generating a spatially variable kernel according to an embodiment.
10 is a diagram referenced to describe a method of generating a spatially variable kernel according to another embodiment.
11 is a diagram referenced to describe an expanded convolution according to an embodiment.
12 is a diagram referenced to describe a method of applying a spatially variable kernel to a first image according to an embodiment.
13 is a flowchart illustrating a method of operating an image processing apparatus according to an exemplary embodiment.
14 is a block diagram illustrating a configuration of an image processing apparatus according to an exemplary embodiment.

Terms used in this specification will be briefly described, and the present invention will be described in detail.

The terms used in the present invention have been selected as currently widely used general terms as possible while considering the functions in the present invention, but these may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

In the entire specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated. In addition, terms such as "...unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software, or a combination of hardware and software. .

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art to which the present invention pertains can easily implement them. However, the present invention may be embodied in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

1 is a diagram illustrating an operation of an image processing apparatus processing an image using an image processing network according to an exemplary embodiment.

Referring to FIG. 1 , an image processing network 30 according to an embodiment receives a first image 10 and processes the first image 10 to generate a second image 20 with improved image quality. It may be a network that In this case, the first image 10 may be an image including noise or a low-resolution image. The image processing apparatus 100 uses the image processing network 30 to perform detailed edge processing of the first image 10 and denoising to remove noise while maintaining the texture, thereby performing the second An image 20 may be generated. The second image 20 may be a higher-resolution image than the first image 10 , and may be an image with improved image quality compared to the first image 10 .

The image processing performed by the image processing network 30 according to an embodiment will be described in detail with reference to the drawings below.

2 is a diagram illustrating an image processing network according to an exemplary embodiment.

Referring to FIG. 2 , an image processing network 30 according to an embodiment includes a similarity calculator 210 , a weight map generator 220 , a spatially variable kernel generator 230 , and a filter unit 240 . can do.

The image processing network 30 according to an embodiment may include a structure for receiving the first image 10 and outputting the second image 20 .

The similarity calculator 210 according to an embodiment may generate similarity information between each of the pixels included in the first image 10 and a neighboring pixel. The similarity information may be information indicating a difference between a pixel value of each of the pixels and a pixel value of a neighboring pixel located around each of the pixels. A method of generating similarity information will be described in detail with reference to FIG. 3 .

3 is a diagram referenced to describe a method of generating similarity information according to an embodiment.

The similarity calculator 210 according to an embodiment may generate similarity information by calculating a difference between each of the pixels included in the first image 10 and a neighboring pixel. For convenience of description, it is assumed that the width of the first image 10 is W, the height is H, and the number of channels is 1 in the embodiments of the present disclosure.

Referring to FIG. 3 , the similarity calculator 210 includes K ² pixels included in the first area 301 centered on the first pixel 310 among a plurality of pixels included in the first image 10 . A difference value between each of the pixels and the first pixel 310 may be calculated. In this case, K x K, which is the size of the first region 301 , may be determined based on the size of a spatial kernel, which will be described later.

The similarity calculator 210 calculates a difference value between each of the K ² pixels included in the first area 301 and the first pixel 310 , thereby generating K ² difference values for the first pixel 310 . can be obtained For example, as shown in FIG. 3 , the similarity calculator 210 calculates a difference value between the first pixel 310 and the first neighboring pixel 311 , the first pixel 310 and the second neighboring pixel 311 . 312 , a difference value between the first pixel 310 and the third neighboring pixel 313 , and a difference value between the first pixel 310 and the fourth neighboring pixel 314 may be calculated. The similarity calculator 210 may obtain K ² difference values for each of the pixels included in the first image 10 other than the first pixel 310 in the same manner. For example, the similarity calculator 210 may obtain K ² difference values from neighboring pixels by using each of the pixels other than the first pixel 310 as a central pixel.

The similarity calculator 210 may arrange K ² difference values for each of the pixels in the channel direction of the corresponding pixel in the similarity information 350 , and accordingly, The size may be W x H, and the number of channels may be K ² .

The first channel image of the similarity information 350 according to an embodiment includes each of the pixels included in the first image 10 and a neighboring pixel (eg, each of the pixels) having a first relative position with respect to each of the pixels. (a pixel at a position moved to the left by (K-1)/2 pixels and upward by (K-1)/2 pixels) based on . In addition, the second channel image of the similarity information 350 includes each of the pixels included in the first image 10 and a neighboring pixel (eg, each of the pixels based on a second relative position). (a pixel at a position moved left by (K-1)/2 - 1 pixel and upward by (K-1)/2 pixels) can be expressed as a difference value. However, the present invention is not limited thereto.

A method by which the similarity calculator 210 obtains K ² difference values for each of the pixels according to an exemplary embodiment will be described in detail with reference to FIGS. 4 and 5 .

4 and 5 are diagrams referenced to describe methods of obtaining similarity information according to an embodiment.

Referring to FIG. 4 , the similarity calculator 210 according to an exemplary embodiment calculates each of the pixels included in the first image 10 in the horizontal direction as p pixels (-(K-1)/2≤p≤(K). K ² images ( 410) can be obtained. In this case, each of the K ² images 410 has the same size (W x H) as the first image 10 .

The similarity calculator 210 may obtain similarity information 350 by calculating a difference image between each of the K ^two images 410 and the first image 10 . Accordingly, the similarity information 350 may have a size of W x H and K ² channels as described with reference to FIG. 3 .

Referring to FIG. 5 , the similarity calculator 210 according to an embodiment may obtain similarity information by performing mask processing on the first image 10 .

The mask processing may be performed through a convolution operation of the first image 10 and each of the mask filters M1, M2, M3, ..., Mn. In this case, n may be K ² −1, and K ² −1 channel images 521 , 522 , 523 , included in the similarity information 350 through mask processing using K ² −1 mask filters. .., 529) can be created. For example, the similarity calculator 210 generates a first channel image 521 of the similarity information 350 through a convolution operation between the first image 10 and the first mask filter M1, The second channel image 522 of the similarity information 350 may be generated through a convolution operation between the first image 10 and the second mask filter M2 . In addition, a third channel image 523 of the similarity information 350 is generated through a convolution operation between the first image 10 and the third marks filter M3, and the first image 10 and the n-th mask Through a convolution operation with the filter Mn, the (K ² −1)th channel image 529 of the similarity information 350 may be generated.

Referring to FIG. 5 , the similarity calculator 210 calculates K×K pixel values included in the first region 501 of the first image 10 and K×K pixel values included in each of the mask filters ( The pixel values included in the similarity information 350 may be calculated by multiplying and summing the parameter values.

In this case, parameter values included in the mask filter may be determined according to positions of neighboring pixels for calculating the similarity information 350 . For example, the first mask filter M1 may have a center pixel, a center pixel and a first relative position (eg, (K-1)/2 pixels to the left with respect to the reference pixel, and (K) It may be a mask filter for calculating similarity information of neighboring pixels having a position moved by -1)/2 pixels). Accordingly, in the first mask filter M1 , a central pixel value may be '1', a pixel value having a first relative position with the central pixel may be '-1', and the remaining pixel values may be '0'.

The similarity calculator 210 according to an embodiment performs a convolution operation between the first region 501 centered on the first pixel 515 and the first mask filter M1 to obtain the similarity information 350 . A value of the second pixel 531 included in the first channel image 521 may be calculated. In this case, the position of the first pixel 515 in the first image 10 may be the same as the position of the second pixel 531 in the first channel image 521 of the similarity information 350 . The value of the second pixel 531 included in the first channel image 521 is obtained by subtracting the value of the first pixel 515 and the pixel 511 having a first relative position from the value of the first pixel 515 . can be a value.

According to a convolution operation using the first mask filter M1 , the first mask filter M1 slides in the horizontal direction and the vertical direction, and each of the pixels included in the first image 10 becomes the first mask filter It may be located at the center of (M1). In this case, the reference pixel is positioned at the center of the area covered by the first mask filter M1 sliding and moving on the input image 10 . The similarity calculator 210 may calculate pixel values included in the first channel image 521 by performing a convolution operation between the changed region and the first mask filter M1 .

In addition, the second mask filter M2 includes a center pixel, a center pixel and a second relative position (eg, (K-1)/2 - 1 pixels to the left with respect to the reference pixel, and an upper (K) It may be a mask filter for calculating similarity information of neighboring pixels having a position moved by -1)/2 pixels). Accordingly, in the second mask filter M2, a central pixel value may be '1', a pixel value having a second relative position with the central pixel may be '-1', and the remaining pixel values may be '0'.

The similarity calculator 210 performs a convolution operation between the first region 501 centered on the first pixel 515 and the second mask filter M2, thereby generating the second channel image ( A value of the third pixel 532 included in 522 may be calculated. In this case, the position of the first pixel 515 in the first image 10 and the position of the third pixel 532 in the second channel image 522 may be the same. Accordingly, the value of the third pixel 532 included in the second channel image 522 is the value of the pixel 512 having a second relative position with the first pixel 515 in the value of the first pixel 515 . It may be a value minus a value.

In the same manner, the region to be subjected to the convolution operation is changed so that each of the pixels included in the first image 10 is positioned at the center of the region to be subjected to the convolution operation, and the changed region and the second mask are changed. By performing a convolution operation with the filter M2, pixel values included in the second channel image 522 may be calculated.

In addition, the third mask filter M3 may be a mask filter for calculating similarity information between the central pixel and the central pixel and peripheral pixels having a third relative position, and the nth mask filter Mn includes the central pixel, It may be a mask filter for calculating similarity information between the central pixel and the neighboring pixels having an n-th relative position.

As illustrated and described with reference to FIG. 5 , the similarity calculator 210 according to an exemplary embodiment performs mask processing using K ² −1 mask filters, so that each of the pixels included in the first image 10 is and similarity information including difference values from neighboring pixels having first to K ² −1 th relative positions with respect to each of the pixels. For example, the similarity calculator 210 uses the first to K ² −1 th masks M1 , M2 , M3 ,,,, Mn to obtain the first to K ² th - One-channel images 521 , 522 , 523 , ..., 529 may be generated.

Also, the similarity information 350 according to an embodiment may include a K ^- th channel image indicating similarity information with itself for each of the pixels included in the first image 10 . Accordingly, in the K ^- th channel image, all pixel values may be '0'.

Meanwhile, the methods of obtaining the similarity information 350 shown and described with reference to FIGS. 4 and 5 are only examples, and the similarity calculator 210 uses various methods to calculate each of the pixels included in the first image 10 and the surrounding area. It is possible to obtain similarity information with pixels.

Referring back to FIG. 2 , the weighted map generator 220 may generate a weighted map based on similarity information. For example, the image quality of the image-processed image may be improved by using a weight map generated to give a large weight to neighboring pixels having similar pixel values for image processing. Accordingly, the weight map generator 220 may generate a weight map indicating weight information corresponding to each pixel based on similarity information between each pixel included in the first image and neighboring pixels. A method of generating the weight map will be described in detail with reference to FIG. 6 .

6 is a diagram referenced to describe a method of generating a weighted map according to an embodiment.

Referring to FIG. 6 , the weight map generator 220 may generate a weight map 650 using the convolutional neural network 610 . The similarity information 350 generated by the similarity calculator 210 may be input to the convolutional neural network 610 , and by passing through the convolutional neural network 610 , a weight map 650 may be output. In this case, the weight map 650 may have a size of W x H and K ² channels.

The convolutional neural network 610 according to an embodiment may include one or more convolutional layers 620 and one or more activation layers 630 . In this case, each of the activation layers may be located after the convolution layer.

Each of the layers included in the convolutional neural network 610 receives values output from the previous layer, performs an operation in the corresponding layer to obtain result values, and outputs the obtained result values to the next layer. . For example, in the convolution layer 620 , a convolution operation between a kernel included in the convolution layer and values input to the convolution layer 620 may be performed. A convolution operation performed in the convolution layer 620 will be described in detail with reference to FIGS. 7 and 8 .

7 and 8 are diagrams referenced to describe a convolution operation performed in a convolution layer according to an embodiment.

7 illustrates an input image (or an input feature map, F_in) input to the convolutional layer 620, a kernel included in the convolutional layer 620, and a convolutional layer 620, according to an embodiment. It is a diagram illustrating an output image (or an output feature map, F_out) outputted from .

Referring to FIG. 7 , the size of the input image F_in input to the convolutional layer 620 according to an embodiment may be W x H, and the number of channels may be C _in . Also, the convolutional layer 620 may include a kernel, and the kernel may include C _out sub-kernels. Also, one sub-kernel may have a size of Kw x Kh x C _in . The number of channels C _in of one sub-kernel may be the same as the number of channels C _in of the first image F_in. The convolution layer 620 may generate an output image F_out by performing a convolution operation between the input image F_in and the kernel. In this case, the size of the output image F_out may be W x H, and the number of channels of the output image F_out may be determined by the number of sub-kernels C _out of the kernel.

8 is a first channel image 720 of an output image F_out through a convolution operation between the input image F_in and the first sub-kernel 710 included in the kernel, according to an embodiment. It is a drawing referenced to explain the process of this creation.

In FIG. 8 , for convenience of explanation, it is assumed that the input image F_in has a size of 5×5 and the number of channels is 1. In addition, it is assumed that one sub-kernel included in the kernel applied to the input image F_in) has a size of 3 x 3, and the number of channels Cin is 1.

Referring to FIG. 8 , a process of extracting features of the input image F_in by applying the first sub-kernel 710 from the upper left to the lower right of the input image F_in is illustrated. In this case, the first sub-kernel 710 has a size of 3 x 3, and the number of channels is 1. For example, a convolution operation may be performed by applying the first sub-kernel 710 to pixels included in the upper left 3×3 region 821 of the input image F_in.

That is, by multiplying and summing pixel values included in the upper left 3 x 3 region 821 and parameter values included in the first sub-kernel 710, one pixel value mapped to the upper left 3 x 3 region 821 ( 831) can be created.

In addition, pixel values included in the 3 x 3 region 822 moved by one pixel from the upper left 3 x 3 region 821 of the input image F_in to the right and parameter values included in the first sub-kernel 710 are added to the input image F_in. By multiplying and summing, one pixel value 832 mapped to the 3×3 area 822 may be generated.

In the same manner, parameter values included in the first sub-kernel 710 and the input image ( By multiplying and summing pixel values of F_in), pixel values included in the first channel image 720 of the output image F_out may be generated. In this case, the data to be subjected to the convolution operation may be sampled while moving one pixel at a time, or may be sampled while moving by the number of two or more pixels. A size of an interval between pixels sampled in the sampling process is referred to as a stride, and the size of the output image F_out may be determined according to the size of the stride. Also, as shown in FIG. 8 , padding may be performed to make the size of the output image F_out equal to that of the input image F_in. The padding increases the size of the input image F_in by giving a specific value (eg, '0') to the edge of the input image F_in to prevent the size of the output image F_out from being reduced. means to do If a convolution operation is performed after padding, the size of the output image F_out may be the same as the size of the input image F_in. However, the present invention is not limited thereto.

Meanwhile, although FIG. 8 shows only the result of the convolution operation for the first sub-kernel 710 (the first channel image 720 of the output image F_out), the convolution operation is performed on Cout sub-kernels. In this case, an output image F_out including Cout number of channel images may be output. That is, the number of channels of the output image F_out may be determined according to the number of sub-kernels included in the kernel.

Again, referring to FIG. 6 , an activation layer 630 may be positioned after the convolution layer 620 .

In the activation layer 630 according to an embodiment, an activation function operation for applying an activation function to values input to the activation layer 630 may be performed. The activation function calculation is to give a non-linear characteristic to the first characteristic information, and the activation function is a sigmoid function, a Tanh function, a Rectified Linear Unit (ReLU) function, a leaky ReLu function, etc. may include, but is not limited thereto.

Also, the convolutional neural network 610 may further include a layer 640 that performs an element-by-element summation operation.

The element-by-element summing operation is an operation of adding values at the same position when each of the values included in the first input information input to the element-by-element summing layer 640 and each of the values included in the second input information are summed.

Accordingly, the similarity information 350 according to an embodiment includes one or more convolutional layers 620 , one or more activation layers 630 , and one or more element-specific summing layers included in the convolutional neural network 610 . By passing through 640 , a weight map 650 may be output.

Referring again to FIG. 2 , the spatially variable kernel generating unit 230 according to an embodiment generates a spatially variable kernel based on the weighted map 650 and the spatial kernel generated by the weighted map generating unit 220 . can In this case, the spatial kernel represents weight information according to a positional relationship between each of the pixels included in the first image 10 and a neighboring pixel. A method of generating a spatially variable kernel will be described in detail with reference to FIG. 9 .

9 is a diagram referenced to describe a method of generating a spatially variable kernel according to an embodiment.

Referring to FIG. 9 , the spatially variable kernel generator 230 may generate the spatially variable kernel 950 using the spatial kernel 910 and the weighted map 650 . For example, the spatially variable kernel generator 230 may convert the spatial kernel 910 into a one-dimensional vector 920 . The spatial kernel 910 has a size of K x K, and pixel values included in the spatial kernel 910 have the largest central pixel value, and the pixel value decreases as the distance from the central pixel increases. The spatially variable kernel generator 230 may arrange the pixel values included in the spatial kernel 910 in the channel direction, and may convert it into a weight vector 920 having a size of 1 x 1 x K ² .

Meanwhile, the weight map 650 generated by the weight map generator 220 may have a size of W x H and K ² channels.

The spatially variable kernel generator 230 may generate the spatially variable kernel 950 by multiplying the weight map 650 and the weight vector 920 . At this time, the spatially variable kernel generator 230 includes each of the one-dimensional vectors having a size of 1 x 1 x K ² and a weight vector 920 having a size of 1 x 1 x K ² included in the weight map 650 . By performing element-wise multiplication of , the spatially variable kernel 950 may be generated.

As shown in FIG. 9 , a spatially variable kernel by performing a multiplication operation for each element of the first vector 651 having a size of 1 x 1 x K ² included in the weight map 650 and the weight vector 920 . A second vector 951 included in 950 may be generated. In this case, the position of the first vector 651 in the weight map 650 and the position of the second vector 951 in the spatially variable kernel 950 may correspond to each other.

The spatially variable kernel 950 according to an embodiment may have the same size as the weight map 650 , W x H, and the number of channels may be K ² .

10 is a diagram referenced to describe a method of generating a spatially variable kernel according to another embodiment.

Referring to FIG. 10 , the spatially variable kernel generator 230 performs a dilated convolution operation between the first weight map 1010 and the first kernel 1020 , and thus the second weight map 1030 is ) can be created. In this case, the first weight map 1010 may have a configuration corresponding to the weight map 650 of FIG. 9 .

The expanded convolution operation means that the convolution operation is performed by applying the kernel to an area larger than the size of the kernel. The expanded convolution operation will be described in detail with reference to FIG. 11 .

11 is a diagram referenced to describe an expanded convolution according to an embodiment.

In FIG. 11 , for convenience of explanation, it is assumed that the input image 1110 , which is the target of the expanded convolution operation, has a size of 7×7 and the size of the kernel 1120 is 3×3.

In the case of a general convolution operation, the kernel 1120 may be applied to a region having a size of 3×3 of the input image 1110 . For example, by multiplying and summing pixel values included in the first region 1131 having a size of 3×3 of the input image 1110 and parameter values of 3×3 included in the kernel, the output image 1140 is Produces one pixel 1145 value.

On the other hand, in the case of the expanded convolution operation, the size of the region to which the kernel 1120 is applied may be expanded according to a delation rate. When the expansion rate is 2, the size of the area to which the kernel is applied may be expanded from 3 x 3 to 5 x 5. For example, as shown in FIG. 11 , the kernel 1120 may be applied to the second area 1132 having a size of 5×5. At this time, by multiplying and summing the values of the first to ninth pixels (shaded pixels) included in the second region 1120 by 9 parameter values included in the kernel 1120 , the pixel 1145 of the output image 1140 . ) value can be generated.

Referring back to FIG. 10 , the spatially variable kernel generator 230 according to an embodiment performs a depthwise dilated convolution between corresponding channels of the first weight map 1010 and the first kernel 1020 . By performing an operation, the second weight map 1030 may be generated.

The size of the first weight map 1010 according to an embodiment may be W x H, and the number of channels may be K ² . Also, the number of channels of the first kernel 1020 may be the same as the number of channels of the first weight map 1010 K ² . Accordingly, the first kernel 1020 may include K ² sub-kernels in the channel direction.

The spatially variable kernel generating unit 230 performs an expanded convolution operation between one channel image included in the first weight map 1010 and one sub-kernel corresponding to the channel image, to thereby perform an expanded convolution operation on the second weight map. One channel image included in 1030 may be generated. For example, the spatially variable kernel generator 230 performs an expanded convolution operation between the first channel image of the first weight map 1010 and the first sub-kernel 1021 included in the first kernel 1020 . Thus, the first channel image of the second weighted map 1030 may be generated.

The spatially variable kernel generator 230 performs an expanded convolution operation on each of the ^second to K2 channel images and the ^second to K2 sub-kernels of the first weight map 1010 in the same manner as described above. By performing , ^second to K-th channel images included in the second weight map 1030 may be generated.

The spatially variable kernel generator 230 converts the spatial kernel 910 into a weight vector 920 having a size of 1 x 1 x K ² , and multiplies the second weight map 1030 by the weight vector 920 . , a spatially variable kernel 1050 may be generated. In this case, the spatially variable kernel generating unit 230 includes each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the second weight map 1030 and a weight vector having a size of 1 x 1 x K ² ( By performing elementwise multiplication of 920 , the spatially variable kernel 1050 may be generated. Since this has been described in detail with reference to FIG. 9 , the same description will be omitted.

Referring back to FIG. 2 , the spatially variable kernel generation unit 230 outputs the generated spatially variable kernels 950 and 1050 to the filter unit 240 , and the filter unit 240 inputs the first image 10 . In response, the second image 20 may be generated by applying the spatially variable kernels 950 and 1050 to the first image 10 . A method of generating the second image 20 by applying the spatially variable kernels 950 and 1050 to the first image 10 will be described in detail with reference to FIG. 12 .

12 is a diagram referenced to describe a method of applying a spatially variable kernel to a first image according to an embodiment.

Referring to FIG. 12 , the spatially variable kernel 1150 according to an embodiment may be the spatially variable kernel 950 of FIG. 9 or the spatially variable kernel 1050 of FIG. 10 , but is not limited thereto.

The spatially variable kernel 1150 may include a kernel vector corresponding to each of the pixels included in the first image 10 . For example, the spatially variable kernel 1150 may include a first kernel vector 1151 corresponding to the first pixel 1210 included in the first image 10 , and is included in the first image 10 . A second kernel vector 1152 corresponding to the second pixel 1220 may be included. Also, the spatially variable kernel 1150 may include a third kernel vector 1153 corresponding to the third pixel 1230 included in the first image 10 .

The filter unit 240 may convert a one-dimensional kernel vector having a size of 1 x 1 x K ² into a two-dimensional kernel having a size of K x K. For example, the first kernel vector 1151 is the first kernel 1215 , the second kernel vector 1152 is the second kernel 1225 , and the third kernel vector 1153 is the third kernel 1235 . can be converted to

The filter unit 240 filters the second image 20 by applying the first kernel 1215 to the first region centered on the first pixel 1210 included in the first image 10 . A value of the fourth pixel 1240 may be calculated. In addition, the filter unit 240 performs filtering by applying the second kernel 1225 to a second region centered on the second pixel 1220 included in the first image 10 , and thus the second image 20 ) of the fifth pixel 1250 may be calculated. In addition, the filter unit 240 performs filtering by applying the third kernel 1235 to a third region centered on the third pixel 1230 included in the first image 10 , and thus the second image 20 ) of the sixth pixel 1260 may be calculated.

In the same way, the filter unit 240 applies a kernel corresponding to each of the pixels included in the first image 10 to an area centered on each of the pixels included in the first image 10, By performing the filtering, pixel values included in the second image 20 may be calculated.

13 is a flowchart illustrating a method of operating an image processing apparatus according to an exemplary embodiment.

Referring to FIG. 13 , the image processing apparatus 100 according to an embodiment may obtain similarity information between each of the pixels included in the first image and a neighboring pixel of each of the pixels ( S1310 ).

The image processing apparatus 100 according to an embodiment calculates a difference value between each of the K ² pixels and the first pixel included in the first area centered on the first pixel among a plurality of pixels included in the first image. can be calculated The image processing apparatus 100 may obtain K ² difference values for the first pixel, and each of the K ² difference values is a pixel value corresponding to the first pixel in each of the channel images included in the similarity information. can be decided. Accordingly, the size of the similarity information may be the same as that of the first image, and the number of channels may be K ² .

Since the method of generating similarity information according to an embodiment has been described in detail with reference to FIGS. 3 to 5 , a detailed description thereof will be omitted.

The image processing apparatus 100 may generate a weighted map based on the similarity information (S1320).

For example, when the image processing apparatus 100 performs image processing by giving a large weight to neighboring pixels having similar pixel values, the image quality of the image-processed image may be improved. Accordingly, the image processing apparatus 100 may generate a weight map indicating weight information corresponding to each pixel based on similarity information between each pixel included in the first image and neighboring pixels.

The image processing apparatus 100 may input similarity information to the convolutional neural network, and may generate a weighted map by passing the input similarity information through the convolutional neural network. In this case, the convolutional neural network may include one or more convolutional layers, one or more activation layers, and one or more element-specific summation layers. Since the method of generating the weight map has been described in detail with reference to FIGS. 6 to 8 , a detailed description thereof will be omitted.

The image processing apparatus 100 according to an embodiment may generate a spatially variable kernel based on the spatial kernel and the weight map ( S1330 ).

For example, the image processing apparatus 100 may convert a spatial kernel into a one-dimensional vector. The spatial kernel has a size of K x K. Pixel values included in the spatial kernel have the largest central pixel value, and the pixel value decreases as the distance from the central pixel increases. The image processing apparatus 100 may arrange pixel values included in the spatial kernel in the channel direction and convert the pixel values into a weight vector having a size of 1 x 1 x K ² .

The size of the weight map generated in step 1320 ( S1320 ) may be W x H, and the number of channels may be K ² .

The image processing apparatus 100 may generate a spatially variable kernel by multiplying the weight map and the weight vector. In this case, the image processing apparatus 100 performs elementwise multiplication of each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the weight map and a weight vector having a size of 1 x 1 x K ² , A spatially variable kernel can be created.

Since the method of generating the spatially variable kernel has been described in detail with reference to FIGS. 9 to 11 , a detailed description thereof will be omitted.

The image processing apparatus 100 according to an embodiment may generate a second image by applying a spatially variable kernel to the first image (S1340).

The spatially variable kernel generated in operation 1330 ( S1330 ) may include a kernel vector corresponding to each of the pixels included in the first image. For example, the spatially variable kernel may include a first kernel vector corresponding to a first pixel included in the first image, and may include a second kernel vector corresponding to a second pixel included in the first image. have.

The image processing apparatus 100 may convert a one-dimensional kernel vector having a size of 1x1x K ² into a two-dimensional kernel having a size of K x K. For example, the first kernel vector may be converted into a two-dimensional first kernel, and the second kernel vector may be converted into a two-dimensional second kernel.

The image processing apparatus 100 calculates a third pixel value included in the second image by applying the first kernel to the area centered on the first pixel and performs filtering, and applies the first kernel to the area centered on the second pixel. By performing filtering by applying the second kernel, a fourth pixel value included in the second image may be calculated.

Accordingly, when filtering the first image, the image processing apparatus 100 may perform filtering by applying different kernels according to the position of the central pixel.

14 is a block diagram illustrating a configuration of an image processing apparatus according to an exemplary embodiment.

Referring to FIG. 14 , the image processing apparatus 100 according to an embodiment may include a processor 120 and a memory 130 .

The processor 120 according to an embodiment may control the image processing apparatus 100 as a whole. The processor 120 according to an embodiment may execute one or more programs stored in the memory 130 .

The memory 130 according to an embodiment may store various data, programs, or applications for driving and controlling the image processing apparatus 100 . A program stored in the memory 130 may include one or more instructions. A program (one or more instructions) or an application stored in the memory 130 may be executed by the processor 120 .

The processor 120 according to an embodiment may include at least one of a central processing unit (CPU), a graphic processing unit (GPU), and a video processing unit (VPU). Alternatively, according to an embodiment, it may be implemented in the form of a system on chip (SoC) in which at least one of a CPU, a GPU, and a VPU is integrated. Alternatively, the processor 120 may further include a Neural Processing Unit (NPU).

The processor 120 according to an embodiment may generate an output image that has been subjected to detailed edge processing and denoising while maintaining a texture while removing noise from the input image by using the image processing network 30 . For example, the processor 120 operates the similarity calculator 210 , the weight map generator 220 , the spatially variable kernel generator 230 , and the filter unit 240 illustrated and described with reference to FIGS. 2 to 12 . At least one of these may be performed.

The processor 120 may obtain similarity information between each of a plurality of pixels included in the first image and a neighboring pixel of each of the pixels. For example, the processor 120 may obtain the first similarity information based on a difference between each of the pixels included in the first image and a first neighboring pixel having a first relative position with respect to each of the pixels. . Also, the processor 120 may obtain second similarity information based on a difference between each of the pixels included in the first image and a second neighboring pixel having a second relative position with respect to each of the pixels.

The processor 120 may generate a weight map based on the similarity information. The processor 120 inputs similarity information to the convolutional neural network, and the input similarity information passes through the convolutional neural network, so that a weighted map may be output. In this case, the convolutional neural network may include one or more convolutional layers, one or more activation layers, and one or more element-specific summation layers. Since the method of generating the weight map has been described in detail with reference to FIGS. 6 to 8 , a detailed description thereof will be omitted.

The processor 120 may generate a spatially variable kernel based on the spatial kernel and the weight map. For example, the processor 120 may transform the spatial kernel into a one-dimensional vector. The spatial kernel has a size of K x K. Pixel values included in the spatial kernel have the largest central pixel value, and the pixel value decreases as the distance from the central pixel increases. The processor 120 may arrange the pixel values included in the spatial kernel in the channel direction and convert the pixel values into a weight vector having a size of 1 x 1 x K ² . The size of the weight map according to an embodiment may be W x H, the number of channels may be K ² , and the processor 100 may generate a spatially variable kernel by multiplying the weight map and the weight vector. In this case, the processor 120 performs elementwise multiplication of each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the weight map and a weight vector having a size of 1 x 1 x K ² , thereby spatially variable You can create a kernel.

The processor 120 may generate the second image by applying a spatially variable kernel to the first image. The spatially variable kernel may include a kernel vector corresponding to each of the pixels included in the first image. For example, the spatially variable kernel may include a first kernel vector corresponding to a first pixel included in the first image, and may include a second kernel vector corresponding to a second pixel included in the first image. have.

The processor 120 may convert a one-dimensional kernel vector having a size of 1 x 1 x K ² into a two-dimensional kernel having a size of K x K. For example, the first kernel vector may be converted into a two-dimensional first kernel, and the second kernel vector may be converted into a two-dimensional second kernel. The processor 120 calculates a third pixel value included in the second image by performing filtering by applying the first kernel to an area centered on the first pixel, and applies the second kernel to the area centered on the second pixel. By performing filtering by applying the kernel, a fourth pixel value included in the second image may be calculated.

Meanwhile, the image processing network 30 according to an embodiment may be a network trained by a server or an external device. The external device may train the image processing network 30 based on the training data. In this case, the training data may include a plurality of data sets including image data including noise and image data in which an edge characteristic or a texture characteristic is preserved while noise is removed.

The server or external device may determine parameter values included in kernels used in each of a plurality of convolutional layers included in the image processing network 30 . For example, the server or external device is a parameter in the direction of minimizing the difference (loss information) of the image data generated by the image processing network 30 and the image data generated by the image processing network 30 while noise as the training data is removed while the edge characteristics are preserved. values can be determined.

The image processing apparatus 100 according to an embodiment may receive the trained image processing network 30 from a server or an external device and store it in the memory 130 . For example, the memory 130 may store the structure and parameter values of the image processing network 30 according to an embodiment, and the processor 120 uses the parameter values stored in the memory 130, according to an embodiment. It is possible to generate a second image in which edge characteristics are preserved while noise is removed from the first image.

Meanwhile, a block diagram of the image processing apparatus 100 illustrated in FIG. 14 is a block diagram for an exemplary embodiment. Each component of the block diagram may be integrated, added, or omitted according to specifications of the image processing apparatus 100 that are actually implemented. That is, two or more components may be combined into one component, or one component may be subdivided into two or more components as needed. In addition, the function performed in each block is for describing the embodiments, and the specific operation or device does not limit the scope of the present invention.

The method of operating an image processing apparatus according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical 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.

Also, the image processing apparatus and the method of operating the image processing apparatus according to the disclosed embodiments may be included in a computer program product and provided. Computer program products may be traded between sellers and buyers as commodities.

The computer program product may include a S/W program and a computer-readable storage medium in which the S/W program is stored. For example, computer program products may include products (eg, downloadable apps) in the form of S/W programs distributed electronically through manufacturers of electronic devices or electronic markets (eg, Google Play Store, App Store). have. For electronic distribution, at least a portion of the S/W program may be stored in a storage medium or may be temporarily generated. In this case, the storage medium may be a server of a manufacturer, a server of an electronic market, or a storage medium of a relay server temporarily storing a SW program.

The computer program product, in a system consisting of a server and a client device, may include a storage medium of the server or a storage medium of the client device. Alternatively, when there is a third device (eg, a smart phone) that is communicatively connected to the server or the client device, the computer program product may include a storage medium of the third device. Alternatively, the computer program product may include the S/W program itself transmitted from the server to the client device or the third device, or transmitted from the third device to the client device.

In this case, one of the server, the client device and the third device may execute the computer program product to perform the method according to the disclosed embodiments. Alternatively, two or more of a server, a client device, and a third device may execute a computer program product to distribute the method according to the disclosed embodiments.

For example, a server (eg, a cloud server or an artificial intelligence server) may execute a computer program product stored in the server to control a client device communicatively connected with the server to perform the method according to the disclosed embodiments.

Although the embodiments have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concept of the present invention as defined in the following claims are also included in the scope of the present invention. belongs to

Claims (21)

In the image processing apparatus,
a memory storing one or more instructions; and
a processor executing the one or more instructions stored in the memory;
The processor is
obtaining similarity information indicating a similarity between each of the pixels included in the first image and a neighboring pixel of each of the pixels;
generating a weight map including weight information corresponding to each of the pixels based on the similarity information;
Generating a spatially variable kernel including a kernel corresponding to each of the pixels based on the weight map and a spatial kernel including weight information according to the positional relationship between each of the pixels and the neighboring pixels, ,
and generating a second image by applying the spatially variable kernel to the first image. According to claim 1,
The processor is
obtaining first similarity information based on a difference between each of the pixels included in the first image and a first neighboring pixel having a first relative position with respect to each of the pixels;
An image processing apparatus for obtaining second similarity information based on a difference between each of the pixels included in the first image and a second neighboring pixel having a second relative position with respect to each of the pixels. According to claim 1,
The processor is
The image processing apparatus of claim 1, wherein the weight map is generated by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using a convolutional neural network. According to claim 1,
and the number of channels of at least one of the similarity information, the weight map, and the spatially variable kernel is determined based on a size of the spatial kernel. According to claim 1,
The spatial kernel is
A pixel located at the center of the spatial kernel has the largest value, and the pixel value decreases as the distance from the center increases. According to claim 1,
The size of the spatial kernel is K x K, the number of channels of the weight map is K ² ,
The processor is
The pixel values included in the spatial kernel are arranged in a channel direction and converted into a weight vector having a size of 1 x 1 x K ² ,
and generating the spatially variable kernel by performing a multiplication operation on each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the weight map and the weight vector. According to claim 1,
The spatially variable kernel includes a number of kernels equal to the number of pixels included in the first image. 8. The method of claim 7,
The processor is
performing filtering by applying a first kernel included in the spatially variable kernel to a first region centered on a first pixel included in the first image;
The image processing apparatus of claim 1, wherein the second image is generated by applying a second kernel included in the spatial transformation kernel to a second region centered on a second pixel included in the first image and performing filtering. According to claim 1,
The processor is
A first weight map is generated by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using the convolutional neural network,
and generating the weight map by performing a dilated convolution operation between the first weight map and a first kernel. 10. The method of claim 9,
The number of channels in the first weight map is the same as the number of channels in the first kernel;
The processor is
and generating the weight map by performing a depthwise dilated convolution operation between the first weight map and channels corresponding to each other of the first kernel. A method of operating an image processing apparatus, comprising:
obtaining similarity information indicating a similarity between each of the pixels included in the first image and a neighboring pixel of each of the pixels;
generating a weight map including weight information corresponding to each of the pixels based on the similarity information;
Generating a spatially variable kernel including a kernel corresponding to each of the pixels based on the weight map and a spatial kernel including weight information according to a positional relationship between each of the pixels and the neighboring pixels step; and
and generating a second image by applying the spatially variable kernel to the first image. 12. The method of claim 11,
The step of obtaining the similarity information includes:
obtaining first similarity information based on a difference between each of the pixels included in the first image and a first neighboring pixel having a first relative position with respect to each of the pixels; and
and obtaining second similarity information based on a difference between each of the pixels included in the first image and a second neighboring pixel having a second relative position with respect to each of the pixels. method of operation. 12. The method of claim 11,
The step of generating the weight map comprises:
and generating the weight map by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using a convolutional neural network. . 12. The method of claim 11,
and the number of channels of at least one of the similarity information, the weight map, and the spatially variable kernel is determined based on a size of the spatial kernel. 12. The method of claim 11,
The spatial kernel is
A pixel located at the center of the spatial kernel has the largest value, and the pixel value decreases as the distance from the center increases. 12. The method of claim 11,
The size of the spatial kernel is K x K, the number of channels of the weight map is K ² ,
The step of generating the spatially variable kernel comprises:
converting the pixel values included in the spatial kernel in a channel direction into a weight vector having a size of 1 x 1 x K ² ; and
and generating the spatially variable kernel by performing a multiplication operation on each of the one-dimensional vectors having a size of 1 x 1 x K ² included in the weight map and the weight vector. . 12. The method of claim 11,
The spatially variable kernel is
and a number of kernels equal to the number of pixels included in the first image. 18. The method of claim 17,
The generating of the second image comprises:
performing filtering by applying a first kernel included in the spatially variable kernel to a first region centered on a first pixel included in the first image; and
and performing filtering by applying a second kernel included in the spatial transformation kernel to a second region centered on a second pixel included in the first image. 12. The method of claim 11,
The step of generating the weight map comprises:
generating a first weight map by performing a convolution operation between the similarity information and one or more kernels included in the convolutional neural network using the convolutional neural network; and
and generating the weight map by performing a dilated convolution operation between the first weight map and a first kernel. 20. The method of claim 19,
The number of channels included in the first weight map is the same as the number of channels included in the first kernel;
The step of generating the weight map comprises:
and generating the weight map by performing a depthwise dilated convolution operation between the first weight map and channels corresponding to each other of the first kernel. One or more computer-readable recording media in which a program for performing the method of claim 11 is stored.

Images