WO2019143025A1 - Image processing method and device using line input and output - Google Patents

Abstract

Disclosed are an image processing method and device using a linewise operation. The image processing device, according to one embodiment, comprises: a receiver for receiving two-dimensional pixel information; at least one first line buffer for outputting the two-dimensional pixel information into a linewise data line; an operator for performing a convolution operation on the basis of the data line data; and at least one second line buffer for outputting the data line on which the operation has been performed.

Description

Image processing method and apparatus using line input and output

The embodiments described below relate to an image processing method and apparatus using line unit operations.

UHD video is becoming popular in ultra High Definition (UHD) TV, Internet Protocol Television (IPTV) services and smart phone applications. While many high-end TVs and smartphones support 4K UHD video, there are many video streams with full high definition (FHD) resolution (1920x1080) due to legacy acquisition devices and services.

Therefore, a sophisticated upscaling technique that can convert low-resolution (LR) content into high-resolution (HR) is essential. In particular, sophisticated upscaling technology is more important for video upscaling from 2K FHD to 4K UHD.

The upscaling method is classified into two types. A single image upscaling method utilizes local spatial correlation within an LR image to recover lost high frequency detail. On the other hand, video upscaling methods use additional data dimensions (time) to improve performance and are expensive to compute.

A single image up-scaling algorithm can be divided into an interpolation method and a super resolution (SR) method. The interpolation method uses a simple interpolation kernel such as a bilinear or bicubic kernel.

The SR method showed better performance than the previous interpolation method. The basic concept of a learning-based approach is learning the mapping function from LR to HR image or video. Learning-based methods can be classified into two types.

The first is to learn LR-HR mapping (based on the input image internal information) using the surrounding information of the LR image itself and the second to use an external LR-HR image pair (based on external training image or dictionary) .

Machine learning algorithms such as sparse coding, anchored neighbors and linear mapping kernels have been proposed for SR.

However, since this learning-based SR method requires a large capacity frame buffer to store intermediate images in most cases, a high-resolution image in the low-complexity hardware (HW) There is a difficulty in implementing the SR method for real-time conversion from resolution to resolution, for example, FHD to 4K UHD.

In recent years, Deep Neural Network (DNN), especially Convolutional Neural Networks (CNNs), has shown excellent performance in various computer vision fields such as image classification, object detection and segmentation.

Machine learning-based methods find features in hand-crafted designs and use these handcrafted features to learn mappings, but in-depth neural networks learn the best features and mappings themselves, And effectively.

Sophisticated CNN-based SR methods have been proposed to improve the visual quality of HR reconstruction. This CNN structure includes a number of layers and nonlinear functions, and is designed to perform SR and generate HR images or high quality video.

Due to excessive multiplications and other computations, conventional CNNs are known to be difficult to implement in low complexity hardware for real-time applications. Moreover, the analysis of computational complexity and execution time for sophisticated CNN-based SR methods was performed at the level of software (Software (SW)) of CPU (Central Processing Unit) and / or GPU (Graphics Processing Unit) platform. In addition, when the CNN structure is implemented in SW and HW, it is difficult to implement in real time because a plurality of frame buffers must be used to store intermediate feature maps.

Embodiments can provide techniques for processing images using line-by-line operations.

An image processing apparatus according to an embodiment of the present invention includes a receiver for receiving two-dimensional pixel information, at least one first line buffer for outputting the two-dimensional pixel information as a data line on a line-by-line basis, An arithmetic unit for performing the convolution operation, and at least one second line buffer for outputting the data line on which the operation is performed.

The receiver may include a first pixel information converter for converting the first color data of the two-dimensional pixel information into second color data.

The image processing apparatus may further include a second pixel information converter for converting the second color data of the data line on which the operation is performed to the first color data.

A first convolution operator for generating a feature map by performing a convolution operation on the basis of the data line; and a second convolution operator for storing the feature map in units of at least one line, A feature map processor for processing to output in a two-dimensional form, and a second convolution operator for performing a convolution operation based on the feature map output in a two-dimensional form.

The first convolution operator may be implemented in a residual block for learning and outputting a residual signal.

The first convolution operator may include at least one 1-D convolution operator for performing a 1-D (1-dimensional) convolution operation.

The first convolution operator may include a depth-wise convolution operator and a point-wise convolution operator directly connected to the depth-wise convolution operator.

The feature map processor may include a compressor for compressing the feature map in at least one line unit.

The feature map processor may further include at least one third line buffer for storing feature maps compressed on a line-by-line basis.

The feature map processor may further include a decompressor for reconstructing the feature map compressed in units of lines into a two-dimensional feature map.

The second convolution operator may include at least one 2-D convolution operator for performing a 2-D (2-dimensional) convolution operation.

The second convolution operator may include a depth-wise convolution operator and a point-wise convolution operator directly connected to the depth-wise convolution operator.

The calculator may further include a quantizer that quantizes the result of at least one convolution operation.

The calculator may further include a weight buffer for storing parameters used for the convolution operation.

The image processing method according to one embodiment includes the steps of receiving two-dimensional pixel information, outputting the two-dimensional pixel information as at least one data line on a line basis, And outputting the data line on which the operation is performed.

The receiving may include converting the first color data of the two-dimensional pixel information into second color data.

And converting the second color data of the data line on which the operation is performed to the first color data.

Wherein the step of performing includes a step of generating a feature map by performing a first convolution operation based on the data line, a step of storing the feature map in units of lines, Processing the feature map to output in the form of a feature map, and performing a second convolution operation based on the feature map output in a two-dimensional form.

The first convolution operation may be performed in a residual block for learning and outputting a residual signal.

The generating may comprise performing at least one 1-D (1-dimensional) convolution operation.

The generating step may include performing a depth-wise convolution operation and directly performing a point-wise convolution operation on the result of the depth-wise convolution operation.

The processing may include compressing the feature map in units of at least one line.

The processing may further include storing a feature map compressed in at least one line unit.

The processing may further include restoring the feature map compressed in the at least one line unit into a two-dimensional feature map.

The performing the second convolution operation may include performing at least one 2-dimensional (2-D) convolution operation.

The performing the second convolution operation may include performing a depth-wise convolution operation and directly performing a point-wise convolution operation on the result of the depthwise convolution operation. have.

The performing the convolution operation may further include quantizing the result of at least one convolution operation.

The step of performing the convolution operation may further include storing a parameter used in the convolution operation.

Figure 1 shows a schematic block diagram of an image processing apparatus according to one embodiment.

Fig. 2 shows a schematic block diagram of the controller shown in Fig.

Fig. 3 shows a schematic block diagram of the operator shown in Fig. 2. Fig.

FIG. 4 shows a schematic block diagram of the first convolution operator shown in FIG.

5A shows an example of a conventional depth wise separable convolution.

5B shows an example of the operation of the first convolution operator shown in FIG.

FIG. 5C shows another example of the operation of the first convolution operator shown in FIG.

Figure 6 shows a schematic block diagram of the feature map processor shown in Figure 3;

7A shows an example of the operation of the feature map processor shown in FIG.

Fig. 7B shows an example of the compression algorithm of the compressor shown in Fig.

Fig. 8 shows a schematic block diagram of the second convolution operator shown in Fig. 3. Fig.

Fig. 9 shows an example of the hardware structure of the image processing apparatus shown in Fig.

Fig. 10 shows an example of a neural network structure used by the image processing apparatus shown in Fig.

Fig. 11 shows an example of a framework for verifying the image processing apparatus shown in Fig.

12A shows an example of the performance of the image processing apparatus shown in FIG.

12B shows another example of the performance of the image processing apparatus shown in FIG.

Fig. 12C shows another example of the performance of the image processing apparatus shown in Fig.

13A shows an example of an original high resolution image.

13B shows an example of an image processed through the bicubic method.

13C shows an example of an image processed through the SRCNN method.

13D shows an example of an image processed through the SRCNN-Ex method.

13E shows an example of an image processed through the FSRCNN method.

13F shows an example of an image processed through the FSRCNN-s method.

13g shows an example of an image processed through the FSRCNN-s method.

13H shows an example of an image processed by applying the quantized weight to the image processing apparatus shown in FIG.

FIG. 13I shows an example of an image processed by applying the quantized weight and activation by the image processing apparatus shown in FIG. 1. FIG.

13J shows an example of an image processed by applying the quantized weight and activation and intermediate feature map compression to the image processing apparatus shown in FIG.

14A shows another example of the original high-resolution image.

14B shows another example of the image processed through the bicubic method.

14C shows another example of an image processed through the SRCNN method.

14D shows another example of an image processed through the SRCNN-Ex method.

14E shows another example of an image processed through the FSRCNN method.

14F shows another example of an image processed through the FSRCNN-s method.

Figure 14g shows another example of an image processed through the FSRCNN-s method.

14H shows another example of the image processed by applying the quantized weight to the image processing apparatus shown in FIG.

Fig. 14I shows another example of an image processed by applying the quantized weight and activation by the image processing apparatus shown in Fig.

FIG. 14J shows another example of an image processed by applying the quantized weight and activation and intermediate feature map compression to the image processing apparatus shown in FIG. 1. FIG.

15 shows an example in which SR hardware is implemented as an FPGA.

In the following, embodiments will be described in detail with reference to the accompanying drawings. However, various modifications may be made in the embodiments, and the scope of the patent application is not limited or limited by these embodiments. It is to be understood that all changes, equivalents, and alternatives to the embodiments are included in the scope of the right.

The terms used in the examples are used for descriptive purposes only and are not to be construed as limiting. The singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, the terms "comprises" or "having" and the like refer to the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, But do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, or combinations thereof.

The terms first, second, or the like may be used to describe various elements, but the elements should not be limited by terms. The terms may be named for the purpose of distinguishing one element from another, for example, without departing from the scope of the right according to the concept of the embodiment, the first element being referred to as the second element, The second component may also be referred to as a first component.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this embodiment belongs. Terms such as those defined in commonly used dictionaries are to be interpreted as having a meaning consistent with the contextual meaning of the related art and are to be interpreted as either ideal or overly formal in the sense of the present application Do not.

In the following description of the present invention with reference to the accompanying drawings, the same components are denoted by the same reference numerals regardless of the reference numerals, and redundant explanations thereof will be omitted. In the following description of the embodiments, a detailed description of related arts will be omitted if it is determined that the gist of the embodiments may be unnecessarily blurred.

A module in this specification may mean hardware capable of performing the functions and operations according to the respective names described in this specification and may mean computer program codes capable of performing specific functions and operations , Or an electronic recording medium, e.g., a processor or a microprocessor, equipped with computer program code capable of performing certain functions and operations.

In other words, a module may mean a functional and / or structural combination of hardware for carrying out the technical idea of the present invention and / or software for driving the hardware.

Figure 1 shows a schematic block diagram of an image processing apparatus according to one embodiment.

Referring to Figure 1, the image processing apparatus 10 may receive an image and process the received image. For example, the image processing apparatus 10 can process a received image to generate a high-resolution image.

The image processing apparatus 10 can efficiently process images in low specification hardware. The image processing apparatus 10 can learn the neural network based on the received image.

The image processing apparatus 10 can increase the resolution of the image using the learned neural network. For example, the image processing apparatus 10 may perform a super-resolution on a low-resolution image.

The image processing apparatus 10 can provide a hardware-friendly Convolutional Neural Network (CNN) based Super Resolution (SR) method. The image processing apparatus 10 may be implemented on an FPGA to convert a 2K full high definition (FHD) image to a 4K UHD (Ultra High Definition) at 60 fps.

The image processing apparatus 10 can effectively perform the SR using the neural network in hardware with limited computation and memory space. The image processing apparatus 10 can process the LR (Low Resolution) input line by line and keep the parameter value of the convolution filter at a small number. That is, the image processing apparatus 10 can significantly reduce the number of filter parameters compared to the conventional CNN by processing LR (Low Resolution) data in line units.

The image processing apparatus 10 can process the SR image using a cascade of 1-D (1-Dimensional) convolution. The image processing apparatus 10 can save the required line memory by keeping the large acceptance field along the horizontal line while keeping the vertical receptive field small. The line memory may include a line buffer.

The image processing apparatus 10 can combine the residual connection with the depth-wise separable convolution to reduce the number of filter parameters of the neural network and reduce the SR performance .

The image processing apparatus 10 can convert the 32-bit floating point data into the fixed-point data with the fixed-point data without substantially degrading the Peak Signal to Noise Ratio (PSNR) through quantization. In addition, the image processing apparatus 10 can compress the feature map to reduce the line memory required to store the feature map data.

The image processing apparatus 10 includes a receiver 1000 and a controller 2000.

Receiver 1000 may receive an image. An image received by the receiver 1000 may mean an image of an object formed by refraction or reflection of light. For example, an image may include video, pictures, and photographs.

Receiver 1000 may receive an image in the form of pixel information. For example, the receiver 1000 may receive two-dimensional pixel information.

The receiver 1000 may include a first pixel information converter 1100. The first pixel information converter 1100 may convert the first color data of the two-dimensional pixel information into the second color data.

The image received by the first pixel information converter 1100 may include color data. For example, the color data may include RGB channel data and YCbCr channel data.

The first pixel information converter 1100 may convert the first color data into the second color data. The first color data may include RGB channels. The second color data may include a YCbCr channel. For example, the first pixel information converter 1100 may convert the RGB channel of the LR input image into a YCbCr channel.

The image may include a low resolution image and a high resolution image.

The controller 200 can process the received image. The controller 200 can increase the resolution of the received image. The controller 200 may perform super resolution on the received image.

The controller 200 can process an image line by line. The controller 200 may output an image in units of lines and increase the resolution of the image by performing operations on a line-by-line basis.

Fig. 2 shows a schematic block diagram of the controller shown in Fig.

2, the controller 200 includes a first line buffer 2100, an operator 2200, a second pixel information converter 2300, a second line buffer 2400, .

The first line buffer 2100 can output the two-dimensional pixel information received by the receiver 1000 as a data line on a line basis. The first line buffer 2100 may include at least one line buffer. At this time, the number of line buffers can be determined according to the size of the convolution filter.

The calculator 2200 can perform a convolution operation based on the data line. The operation of the calculator 2200 will be described in detail with reference to FIG.

The second pixel information converter 2300 may convert the second color data of the operated data line into the first color data. For example, the second pixel information converter 2300 may convert a YCbCr channel to an RGB channel.

The second line buffer 2400 can output the data line on which the operation is performed. The second line buffer 2400 may include a plurality of line buffers. For example, the second line buffer 2400 may include four output line buffers.

Fig. 3 shows a schematic block diagram of the operator shown in Fig. 2. Fig.

The operator 2200 may include a first convolution operator 2210, a feature map processor 2230, a second convolution operator 2250, a quantizer 2270 and a weight buffer 2290.

The first convolution operator 2210 can generate a feature map by performing a convolution operation based on the data line.

The feature map processor 2230 can store the feature maps in units of at least one line and process the feature maps stored in units of at least one line in a two-dimensional form.

The second convolution operator 2250 can perform a convolution operation based on the feature map output in a two-dimensional form.

The quantizer 2270 can quantize the result of at least one convolution operation. The quantizer 2270 may quantize the convolution result and filter parameters using a variety of quantization methods.

The quantization method of the quantizer 2270 may include any quantization algorithm capable of converting a floating point to a fixed point, including uniform quantization and non-uniform quantization. For example, quantizer 2270 may quantize at least one convolution result and filter parameters using uniform quantization.

The image processing apparatus 10 may use a fixed-point representation via the quantizer 2270 to reduce complexity. The quantizer 2270 can convert a floating-point to fixed-point data.

Fixed-point data can be defined as [IL, FL] to indicate the number. IL (integer length) represents an integer length, and FL (fraction length) represents a fractional (or fractional) length. The quantizer 2270 can calculate the total number of bits used to represent a number by adding a fractional number of bits to the integer number of bits.

The sum of these, IL + FL, can be expressed by the word length WL. Quantizer 2270 may be set to a fixed-point format through at [IL, FL], limits the precision of the FL data bits, and the range ^{[-2 IL-1, 2 IL} -1 -2 -FL] .

The quantizer 2270 can use a rounding method that rounds to nearest when converting a floating point to a fixed point. This rounding method can be expressed by Equation (1).

here,

X is less than or equal to x

Can be defined as the largest integer multiple of. If x is outside the range of [IL, FL], the result can be saturated to the lower or upper limit of [IL, FL]. Finally, the formula for converting a floating-point to a fixed-point can be expressed as shown in Equation 2 below.

The optimal WL, IL, and FL values may be applied to the image processing apparatus 10 to minimize the PSNR degradation of the test set for floating point data (filter parameters and activation values) through a number of experiments. The drop due to quantization in the network used by the image processing apparatus 10 may be very small.

The weight buffer 2290 may store parameters used for the convolution operation.

The above-described components can operate in a pipe-line form. For example, at least one convolution operation may operate in a pipelined fashion.

FIG. 4 shows a schematic block diagram of the first convolution operator shown in FIG.

Referring to FIG. 4, a first convolution operator 2210 may be implemented in a residual block for learning and outputting a residual signal. The first convolution operator 2210 may include at least one 1-D convolution operator for performing a 1-dimensional (1-dimensional) convolution operation.

The 1-D convolution operation may be a convolution operation in which data on a line-by-line basis is input. For example, a 1-D convolution operation may mean 1 x n convolution. At this time, the length n of the line-by-line data may have any integer value of 2 or more.

The first convolution operator 2210 includes a point-wise (PW) convolution operator 2213 directly connected to a depth-wise (DW) convolution operator 2211 and a depthwise convolution operator 2211, . ≪ / RTI >

The depth-wise convolution operator 2211 can perform the convolution operation in the depth direction of the feature map. The depth-wise convolution operator 2211 may perform at least one depthwise convolution operation. The point-wise convolution operator 2213 can perform a convolution operation on a point-by-point basis. The point-wise convolution operator 2213 may perform at least one point-wise convolution operation.

The depth-wise convolution operator 2211 may include at least one 1-D convolution operator. The point-wise convolution operator 2213 can perform 1x1 convolution.

Fig. 5A shows an example of conventional depth wise separable convolution, Fig. 5B shows an example of the operation of the first convolution operator shown in Fig. 3, Fig. 4C shows an example of the operation of the first convolution operator shown in Fig. Another example is shown.

Referring to FIGS. 5A to 5C, the first convolution operator 2210 may perform a convolution operation using depth-wise separable convolutions (DSC). In this way, the first convolution operator 2210 can achieve similar classification performance with only about 1/9 of the number of parameters as compared with the case of the existing non-separable convolution.

The DSC may include a cascaded depthwise convolution operation, a rectified linear unit (ReLU), and a pointwise convolution operation.

DSC can exhibit poor performance when used in regression problems such as SR. Batch normalization (BN) can require relatively high computational complexity to degrade performance and to calculate the mean and variance in regression analysis.

Accordingly, the first convolution operator 2210 can use a structure in which BN and ReLU are removed from the DSC. For example, the first convolution operator 2210 removes the ReLU operation between the depth-wise convolution operator 2211 and the point-wise convolution operator 2213, and can directly use them.

In DSC, when a small number of convolution filters are used with ReLU, the feature map via ReLU may become too sparse. This can interfere with learning and result in lower PSNR performance.

FIG. 5A illustrates a conventional DSC structure, and FIG. 5B illustrates a structure of a convolution layer used by the first convolution operator 220. FIG. The results of PSNR and SSIM performance comparisons between the presence and absence of ReLU between the depth-wise convolution operator 2211 and the point-wise convolution operator 2213 for a Set-5 data set may be as shown in Table 1.

In the conventional DSC, a 3 × 3 size filter can be used for depth-wise convolution. However, some display applications, such as T-Con, may not be able to use line memory excessively, which can limit the use of 3x3 filters in the network. On the other hand, a large acceptance field using a 3x3 or larger filter may be needed to achieve high performance in deep running.

Thus, by using 1-D horizontal convolution for some convolutional layers, the image processing apparatus 10 can make the network more compact and suitable for the hardware in which the LR input data is streamed line by line. For example, the image processing apparatus 10 may apply 1-D horizontal convolution to the first convolution operator 2210.

As a result, the first convolution operator 2210 may have a rectangular vertical field having a longer length in the horizontal direction and a shorter length in the vertical direction. In this way, the image processing apparatus 10 can make the line memory required to store the intermediate feature maps as small as possible.

For an efficient hardware implementation, the network may need to maintain a convolution filter that is as small as possible. However, a small number of filter parameters can worsen the learning of networks including DSC and 1-D horizontal convolution.

This may be because the lack of filter parameters makes the inter-layer connections on the network thin and the restoration of the image is not well learned.

The image processing apparatus 10 can significantly reduce the number of filters while having excellent SR performance by inserting a residual connection in the network. For example, the image processing apparatus 10 may reduce the number of filters by implementing the first convolution operator 2210 in the residual block.

From a hardware perspective, implementing 2-D convolutional residual connections may require additional line memories to store the input of the residual connections, which may also be needed at the end of the connection.

Thus, the image processing apparatus 10 can be easily implemented in hardware by using a delay buffer using only 1-D convolution in the residual connection. Figure 5C may represent the final DSC structure with the final 1-D horizontal convolution and residual connections.

Figure 6 shows a schematic block diagram of the feature map processor shown in Figure 3;

Referring to FIG. 6, the feature map processor 2230 may include a compressor 2231, a third line buffer 2233, and a decompressor 2235.

The compressor 2231 can compress the feature map in units of at least one line. The compressor 2231 can compress the feature map in at least one direction of the width, height, and depth directions of the feature map.

The third line buffer 2233 can store the compressed characteristic map on a line-by-line basis. The third line buffer 2233 may include at least one line buffer. At this time, the number of line buffers can be determined according to the size of a convolution filter used for performing a convolution operation.

The decompressor 2235 can restore the feature maps compressed on a line-by-line basis into a two-dimensional feature map.

Fig. 7A shows an example of the operation of the feature map processor shown in Fig. 3, and Fig. 7B shows an example of the compression algorithm of the compressor shown in Fig.

7A and 7B, the compressor 2231 can compress the feature map through various algorithms. The compression algorithm may include Fixed-Length Coding and Variable-Length Coding. The fixed length coding algorithm may include the DXT algorithm and may include block based algorithms such as JPEG and JPEG 2000. The fixed length coding algorithm can be preferred in terms of hardware complexity.

The variable length coding algorithm may include Huffman and Arithmetic Coding. A variable length coding algorithm can be used to increase the compression rate.

In addition, the compressor 2231 can compress the feature map using a modification algorithm of the above-described algorithm.

The size of the acceptance field can have a significant impact on performance. In addition, both the horizontal and vertical acceptance fields can be important. However, in the case of a 3x3 convolution that includes a vertical direction, the feature map data may have to be stored in the line memory when the data output after passing through the previous convolution layer is transmitted in the next 3x3 convolution layer.

When passing through a 3x3 convolution layer, a line memory that is twice as many as the number of line memories required to store the output feature maps of the current layer may be needed.

The use of many line memories causes many problems in chip design. The main problem is the increase in chip size due to the increase in the number of power rings used in line memory, the routing congestion in place & route (P & R), and the powering on memory memory block boundaries If the power ring is insufficient, it may include the occurrence of a voltage drop.

In order to solve the above-mentioned problem, a method of reducing the number of line memories is required, and a method of compressing the feature map can be considered.

The compressor 2231 can use the feature map compression method considering various points from the hardware implementation viewpoint. The feature map compressed by the compressor 2231 may include an intermediate feature map.

The compressor 2231 can use a very simple compression algorithm. Since compression is considered to reduce the number of line memories, the size of the logic used in the compression method may have to be smaller than the memory size required to store the intermediate feature map before compression.

Since the use of residual learning and ReLU causes many zero and nearly zero values in the feature maps, the compressor 2231 can provide an efficient compression algorithm for this data characteristic. The compressor 2231 can compress the data using only the data adjacent in the horizontal direction in order to use the line memory effectively.

For example, the algorithm used by the compressor 2231 may include an algorithm modified to fit the DXT5 algorithm to the CNN structure. DXT5 is capable of independently compressing each RGB color channel input in compressing RGB pixels illustratively comprised of 4x4 blocks.

The maximum value (MAX) and minimum value (MIN) of each color channel can be calculated. Six intermediate points can be generated through interpolation using the maximum value (MAX) and the minimum value (MIN). MAX, MIN, and 6 intermediate points can be defined as reference colors for compression.

To encode the pixel data, each pixel may be assigned an index value of the color closest to the reference color. The encoding can be completed by storing a 4x4 block index value and MAX and MIN values. There are eight neighboring index values for each pixel in the 4x4 block, and each index can be represented by three bits for each pixel.

The decoding can be easily performed using MAX, MIN values and index values in the reverse order of the encoding process. If bits per pixel (bits per pixel (bpp)) of RGB input is 8 bits, DXT5 can have a fixed compression ratio of 2: 1 for a 4x4 block. The compression ratio (CR) can be calculated as shown in Equation (3).

The compressor 2231 of the image processing apparatus 10 may provide a compression method that modifies the DXT5 algorithm to minimize image quality degradation and increase CR. The difference between the conventional DXT 5 and the image processing apparatus 10 can be shown in Table 2.

The compressor 2231 can fix the minimum value to 0 and calculate only the maximum value. Fixing the minimum value to 0 uses a characteristic that the data of the intermediate feature maps is close to 0 or 0.

The imposition apparatus 10 can reduce the bit for storing the minimum value and eliminate the logic for calculating the minimum value by fixing the minimum value to zero. Since the intermediate feature map data must be processed line by line in hardware, the block size of the feature map data can be set to 1x32.

In addition, a 5-bit index may be assigned to a quantization level for each piece of data in a 1x32 block of the feature map. One index of data is allocated to maintain image quality. The 5 bit length for the indices can be determined experimentally by examining the PSNR performance along the bit length for the data point indices.

The CR of the compressor 2231 can be expressed by Equation (4).

The compression rate may be 2.58: 1 (= 14 (1x32) / (14 + 5 (1x32)) if the word length (word length (WL)) of the feature map data is 14 bit depth after quantization of activation. That is, the number of line memories for storing feature maps can be reduced to about 2.58 times.

As shown in Table 2, the compressor 2231 uses divisor values of 32 (a multiple of 2) instead of 31 (2 ⁵ -1) to reduce hardware complexity for calculating intermediate points. Can be set. With this, the compressor 2231 can calculate the midpoints with shift and add operators (shift and add operators)

Fig. 8 shows a schematic block diagram of the second convolution operator shown in Fig. 3. Fig.

7, the second convolution operator 2250 may include a point-wise convolution operator 2251 and a point-wise convolution operator 2253 directly connected to the depth-wise convolution operator 2251. The second convolution operator 2250 may include at least one 2-D convolution operator for performing a 2-D (2-dimensional) convolution operation.

The depth-wise convolution operator 2251 can perform a convolution operation in the depth direction of the feature map. The depth-wise convolution operator 2251 can perform at least one depthwise convolution operation. The point-wise convolution operator 2253 can perform a convolution operation on a point-by-point basis. The point-wise convolution operator 2253 may perform at least one point-wise convolution operation.

The 2-D convolution operation may be a convolution operation using two-dimensional data as input. For example, a 2-D convolution operation may mean m x n convolution. At this time, the length m and n of the line-by-line data may have an arbitrary integer value of 2 or more.

Although exemplary 3-by-3 convolution is used herein as an example of 2-D convolution, the 2-D convolution used by the image processing apparatus 10 may not be limited thereto.

The depth-wise convolution operator 2251 can perform a convolution operation in the depth direction of the feature map. The point-wise convolution operator 2253 can perform a convolution operation on a point-by-point basis.

The depth-wise convolution operator 2251 may include at least one 2-D convolution operator. The point-wise convolution operator 2253 may perform 1x1 convolution.

Like the first convolution operator 2210, the second convolution operator 2250 can perform the convolution operation using the depthwise separable convolution. The description related to the depthwise separable convolution is the same as the first convolution operator 2210 and will not be described in detail.

Fig. 9 shows an example of the hardware structure of the image processing apparatus shown in Fig.

The image processing apparatus 10 can process a low-resolution image to generate a high-resolution image. For example, the image processing apparatus 10 may generate a 4K UHD image from the FHD image.

The example of FIG. 9 may represent a pipeline hardware architecture for SR. The example of FIG. 9 may be of two types. Type-1 means no compression of the intermediate feature map, and Type-2 means compression.

Details of the image processing apparatus shown in Fig. 9 can be shown in Table 3.

The image processing apparatus 10 includes a first pixel information converter 1100, a first line buffer 2100, a data aligner, a depth-wise convolution operator 2211, a point-wise convolution operator 2213, a compressor 2231 A third line buffer 2233, a decompressor 2235, a depth-wise convolution operator 2251, a point-wise convolution operator 2253, a quantizer 2270, a weight buffer 2290, 2300 and a second line buffer 2400.

The image received by the first pixel information converter 1100 may include color data. For example, the color data may include RGB channel data and YCbCr channel data.

The first pixel information converter 1100 may convert the first color data into the second color data. The first color data may include RGB channels. The second color data may include a YCbCr channel. For example, the first pixel information converter 1100 may convert the RGB channel of the LR input image into a YCbCr channel.

The first line buffer 2100 may illustratively include four line buffers. The depth-wise convolution operator 2211 can perform a 1x5 convolution operation. The third line buffer 2233 may include an even line buffer and an odd line buffer.

The depth-wise convolution operator 2251 can perform a 3x3 convolution operation. The second pixel information information converter 2300 may convert the second color data into the first color data. For example, the second pixel information converter 2300 may convert a YCbCr channel to an RGB channel.

The weight buffer 2290 may store parameters (or filter parameters) used for the convolution operation. The weight buffer 2290 may update parameters received from the convolution operators.

The second line buffer 2400 may include a plurality of line buffers. For example, the second line buffer 2400 may include four output line buffers.

The outputs of all convolution operators may be quantized through a quantizer 2270. All of the weighting parameters may also be quantized through the quantizer 2270.

The quantizer 2270 can convert a 32-bit floating point to a 10-bit fixed point. The weight buffer 2290 may store the quantized weight parameters.

The arrows in FIG. 9 may represent the data path. That is, the arrow may represent a datapath according to type-1 and a datapath according to type-2.

The image processing apparatus 10 can operate in a pipeline structure. The pipeline may refer to a structure in which the output of the processing step of one data is connected to the input of the next step. The processing step of the connected data can be performed in several steps simultaneously or in parallel.

That is, the components included in the image processing apparatus 10 may operate concurrently or in parallel to process images. For example, at least one convolution operation of the image processing apparatus may operate in a pipeline form.

Convolution operators may load the convolution filter parameters from the weight buffer 2290. The first pixel information converter 1100 may then extract the YCbCr value from the RGB input stream. The first line buffer 2100 may store four rows of the YCbCr LR input image for use in the shortest entry upscaling to obtain the interpolated image for the residual connection at the end of the network.

The data aligner may rearrange the data of the four line buffers and input streams of the first line buffer 2100 and generate 3x3 sized YCbCr LR patches. Here, the Y channel of the LR patches can be transmitted in a 3x3 convolution layer.

After the first convolution operation, the feature map may pass through the ReLU activation function. Thereafter, the output of the ReLU function may pass through the first convolution operator 2210. The first convolution operator 2210 may generate a feature map (or an intermediate feature map).

The compressor 2231 compresses the intermediate feature map that has passed through the residual block and the ReLU, and the third line buffer 2233 can store it.

The decompressor 2235 can read the data stored in the third line buffer 2233 and decompress the decompressed data at a timing of one-delayed line data DE (data enable). The depth-wise convolution operator 2251 performs a 3 × 3 convolution operation on the decompressed data, and the point-wise convolution operator 2253 performs a 1 × 1 convolution operation on the output of the depth-wise convolution operator 2251 .

After being output from the point-wise convolution operator 2253, the number of feature map channels may be reduced from 32 to 16 in half. The feature map in which the channel is reduced can again pass through the compressor 2231, the third line buffer 2233, and the decompressor 2235 sequentially. Thereafter, a convolution operation may be performed through the depth-wise convolution operator 2251 and the point-wise convolution operator 2253 once again.

The output after repeated convolution operations can be made up of four channels used to create a 2x2 HR patch in a manner similar to subpixel convolution.

Finally, the image processing apparatus 10 adds the super-resolved Y data (Y _C ) of 2x2 size and the 2X up-sampled data (Y _N ) by the shortest entry interpolation method, (Y _F ).

In order to synchronize the two timings of Y _C and Y _N data, Y _N data is stored in a FIFO (First-In-First-Out) and can be read at the same timing as Y _C. The CbCr data delayed from the FIFO may also be upsampled twice based on the nearest neighbor interpolation and sent to the second pixel information converter to obtain RGB pixels.

The two output buffers of the second line buffer 2400 can store the generated 2x2 RGB HR patches and can be transmitted to the display device every output clock cycle at the output timing.

Four line buffers may be used in the second line buffer 2400 to avoid read / write conflicts with 2x2 RGB HR patches using a double buffering scheme for stream processing.

Fig. 10 shows an example of a neural network structure used by the image processing apparatus shown in Fig.

Referring to FIG. 10, the image processing apparatus 10 can process images using a hardware-friendly CNN-based SR network.

The image processing apparatus 10 can process the image using only a part of the color data. Of the YCbCr channels, only the luminance signal (Y) channel can be input to the CNN network for processing. The performance of learning using only Y channel may be similar to the performance of learning using RGB channel.

When learning a CNN network with RGB channels, the number of parameters used for the RGB channel may be three times greater than the number of parameters used only for the Y channel in the 2-D convolution of the first layer and the point-wise convolution of the last layer .

At this time, the color difference signal (Cb, Cr) channel data can be upscaled using an interpolation method. The interpolation method may include bicubic interpolation and nearest neighbor interpolation.

For example, the image processing apparatus 10 can be upscaled using simple shortest-point interpolation instead of bicubic interpolation for reduced hardware complexity and hardware efficiency. Also, in order to reduce the complexity, the image processing apparatus 10 may learn a neural network using a residual learning technique.

The image processing apparatus 10 can calculate the final HR image Y _F by adding the interpolated LR image Y _N and the output Y _C of the network. This can be expressed by Equation (5).

In order to use the convolution filter parameters and the line memory as little as possible, the image processing apparatus 10 combines depth-wise separable convolutions, 1-D horizontal convolution and residual connection Can be used.

As a result, the number of filter parameters is about 21 times smaller than that of the conventional Super-Resolution Convolutional Neural Network (SRCNN) -Ex, about 4.5 times smaller than that of the FSRCNN (Fast Super Resolution Convolutional Neural Network) While PSNR and Structural Similarity (SSIM) performance may be similar to SRCNN-Ex.

10, the image processing apparatus 10 may perform a convolution operation through a 2-D convolution layer of two layers and a 1-D convolution layer. For example, the 2-D convolution operation may be a 3x3 convolution operation, and the 1-D convolution operation may be a 1x5 convolution operation. In this case, the total receiving field size may be 7 x 15.

Fig. 11 shows an example of a framework for verifying the image processing apparatus shown in Fig.

Referring to FIG. 11, performance may be evaluated for popular data sets to compare bi-cubic and conventional CNN-based SR methods with the image processing apparatus 10. Through the performance evaluation experiments, the performance of the image processing apparatus 10 can be compared with software-based methods including SRCNN, SRCNN-Ex, FSRCNN, FSRCNN-s and VDSR (Very Deep Super-Resolution). Also, the performance of the image processing apparatus 10 can be compared with other real-time SR hardware in view of the gate count and the operating frequency.

A set of benchmark data that are available for learning and testing can be used. In particular, the SR network can be learned using 291 images consisting of 200 images from 91 images from Yang et al. And Berkeley Segmentation Dataset.

Test set-1 and test set-2 can be used for performance comparison. Test set-1 can consist of Set5, Set14, B100 and Urban100, which can often be used as SR benchmarks in many ways. Test Set-2 can consist of 8 4K UHD images and can be used for testing.

All experiments can be performed with a scale factor of 2 for SR. PSNR and SSIM can be used as metrics for evaluation. Since SR was performed on the luminance channel of the YCbCr color space, the PSNR and SSIM can be computed using the Y channel of the reconstructed original HR image.

For learning and testing, the LR input image can be intentionally generated by downsampling from the original HR image using bi-cubic interpolation of double scale. Subimages of size 128 x 128 may be cropped randomly for learning. The LR-HR learning image pair can be augmented using rotation, reflection, and scaling.

The weights may be initialized using a uniform distribution, and the bias may not be used to reduce the number of parameters. L1 loss can be used as a cost function instead of L2 loss. The proposed SR network can be learned using the Adam optimizer.

The learning rate is set to 0.0001 and can be reduced by 10 for every 50 epochs. During the learning, the size of the mini batch can be set to two. A 3.4GHz NVIDIA Titan X GPU and an Intel Core i7-6700 CPU can be used for learning tests.

The weight parameters of the SR network during the learning phase may be quantized according to Equation 2 from floating point to fixed point in the test phase while calculating the floating point.

In addition, only the feature maps of the 3 × 3 depth wise convolution layers can be compressed by quantizing the activation of all the convolution layers and using the compression method. Optimum quantization bits for weighting and activation can be found experimentally and quantized weighting parameters can be used in the image processing apparatus 10.

The compressed intermediate feature map and the final SR image in the algorithm step can be compared with the hardware simulation results designed using the golden model.

FIG. 12A shows an example of the performance of the image processing apparatus shown in FIG. 1, FIG. 12B shows another example of the performance of the image processing apparatus shown in FIG. 1, Another example of performance is shown.

Referring to Figures 12A-12C, weighting parameters and activation may be quantized for hardware implementation. It may be important to find the appropriate quantization bit depth since the quantization for the weight parameter and activation greatly affects the quality of the output image. That is, in the above description, appropriate values for three parameters of the word length (WL), the integer length (IL), and the decimal length (FL) may be required. To do this, experiments can be performed by changing WL, IL, FL for various data sets.

FIG. 12A may show a PSNR graph for WL and IL that quantizes weight parameter values for a Set5 data set. As shown in FIG. 12A, when the bit depth is 10 or more, the PSNR performance of the SR network can be similar to the case of no weighting and active quantization. Also, it can be seen that PSNR performance for 10 bit WL is greatly reduced at IL = 4 or more.

This may have a greater impact on the PSNR performance than the IL bit depth due to the use of the residual network as in Equation (5). For quantization of the weighting parameters, the WL bit depth can be set to 10 bits and the IL bit depth can be set to 2, which can also be used for active quantization and intermediate feature map compression.

FIG. 12B shows the PSNR performance of the SR network for WL and IL bit depths for active quantization. Based on the experimental result of FIG. 12B, WL for active quantization can be set to 14 bits, and IL can be set to 2 bits.

FIG. 12C can show experimental results on a compression method applied to a quantized feature map to reduce line memory usage. Various block sizes and indices (quantization levels) may be examined in relation to PSNR performance. The smaller the number of quantization levels of compression, the higher the compression ratio but the lower the PSNR performance.

Based on the experimental results in FIG. 12C, a block size of 32 bits and an index size of 5 bits (the number of quantization levels) can be selected as a compromise between the required line memory and the resulting PSNR.

13A shows an example of an original high-resolution image, FIG. 13B shows an example of an image processed through a bicubic method, and FIG. 13C shows an example of an image processed through a SRCNN method.

13D shows an example of an image processed through the FSRCNN-s method, FIG. 13E shows an example of an image processed through the FSRCNN method, and FIG. 13F shows an example of an image processed through the FSRCNN- .

FIG. 13G shows an example of an image processed through the VDSR method, and FIG. 13H shows an example of an image processed by applying the quantized weight to the image processing apparatus shown in FIG.

FIG. 13I shows an example of an image processed by applying the quantized weight and activation to the image processing apparatus shown in FIG. 1, FIG. 13J shows an example in which the image processing apparatus shown in FIG. An example of an image processed by applying compression is shown.

FIG. 14A shows another example of the original high-resolution image, FIG. 14B shows another example of the image processed through the bicubic method, and FIG. 14C shows another example of the image processed through the SRCNN method.

14D shows another example of the image processed through the SRCNN-s method, Fig. 14E shows another example of the image processed through the FSRCNN method, and Fig. 14F shows another example of the image processed through the FSRCNN-s method .

FIG. 14G shows another example of the image processed through the FSRCNN-s method, and FIG. 14H shows another example of the image processed by applying the quantized weight to the image processing apparatus shown in FIG.

FIG. 14I shows another example of an image processed by applying the quantized weight and activation to the image processing apparatus shown in FIG. 1, FIG. 14J shows the image processing apparatus shown in FIG. 1 with the quantized weight, Another example of an image processed by applying compression is shown.

13A to 14J, the performance of the image processing apparatus 10 can be compared with other CNN-based SR methods including bicubic and SRCNN, SRCNN-Ex, FSRCNN, and FSRCNN-s. Publicly available MATLAB (TM) source code for SRCNN, SRCNN-Ex, FSRCNN, and FSRCNN-s may be used and the image processing apparatus 10 may be implemented using PyTorch.

For a fair comparison, the boundaries of the HR reconstruction and the original image may be excluded from the PSNR / SSIM calculation. All methods can be executed on the CPU platform.

Because the official code for VDSR is only executable on the GPU platform, third-party code running on the CPU platform can be used to measure PSNR / SSIM and run-time.

The execution time of the image processing apparatus 10 can be measured based on a software implementation using PyTorch.

The average PSNR and SSIM values for the SR method compared to the four benchmark data sets are shown in Table 4.

As shown in Table 4, it can be seen that the image processing apparatus 10 performs better than FSRCNN-s and occupies only 64% of the filter parameters in comparison with FSRCNN-s. In addition, it can be confirmed that the performance of the image processing apparatus 10 does not occur even after the value of the weight parameter and the quantization for activation.

When applying feature map compression to the network of image processing apparatus 10, there is a performance degradation of about 0.1 dB at PSNR, but the space of the required line memory may be reduced by about 2.58 times.

Table 5 shows the performance comparison results of the image processing apparatus 10 and other CNN-based SR methods in terms of the average calculation time of the PSNR and SSIM of the test set 2 composed of the 4K UHD test image.

It can be seen that the image processing apparatus 10 can restore a quality HR image comparable to other SR methods. Regarding execution time, RCNN, SRCNN-Ex, FSRCNN, and FSRCNN-s can use relatively long run times because public code is implemented in MATLAB and may not be optimized on the CPU platform.

The network used by the image processing apparatus 10 for a fair comparison can also be implemented in TensorFlow, and other codes can be created in TensorFlow and the execution time can be measured in the GPU platform.

Referring to Table 5, the execution time measured by the GPU for various CNN-based SR methods including the image processing apparatus 10 can be confirmed. The execution time of the image processing apparatus 10 executed in the GPU was measured to be about 50 ms, which is about three times faster than the FPGA implementation.

13A to 13J can represent restored images and their cropped regions using five CNN-based SR methods including bi-cubic and image processing device 10. [ It can be confirmed that the image processing apparatus 10 uses the smallest number of parameters but the resulting HR image can still be perceived with sharp edges and few artifacts.

FIGS. 14A through 14J may represent cropped regions of the reconstructed HR image for children images of 4K UHD resolution. It can be confirmed that the visual quality of the image processing apparatus 10 and other CNN-based SR methods are similar.

15 shows an example in which SR hardware is implemented as an FPGA.

Referring to FIG. 15, the prototype of the image processing apparatus 10 implemented on the FPGA can be confirmed. Table 6 shows the details of Lee, Yang, previous Super-Interpolation (SI) and implementation of the image processing apparatus 10. [

* Two input NAND gates are counted as one equivalent gate.

Lee et al. Presented hardware using a Lagrange interpolation method using a sharpening algorithm to obtain a 4K UHD video stream from HD and FHD streams at 30 fps.

Yang's HW architecture is based on Anchor Neighborhood Regression (ANR), which uses a dictionary to obtain intermediate images at a target resolution for generating high-frequency patches and obtains FHD at 60 fps.

The machine learning based SI HW architecture can be based on linear mapping using edge orientation analysis, which directly restores the HR image through high frequency reconstruction without the need for intermediate images.

The image processing apparatus 10 may be implemented using a system verilog of the FPGA. The output clock speed of the image processing apparatus 10 may be four times the input clock speed. This may be because the ratio of FHD operating frequencies above 4K UHD is usually 1/4.

The image processing apparatus 10 may process 4 pixels per clock cycle to support a 4K UHD video stream of 60 fps and may be used in combination with the 150MHz target operating frequency and the synthesis and P & R steps of the Vivado Design Suite 2015.4 May be implemented according to the constraints applied.

In addition, Xilinx Kintex UltraScale FPGA KCU105 evaluation board and TED's HDMI 2.0 expansion card can be used to support FHD input and 4K UHD output video interface to validate implemented SR hardware.

The image processing apparatus 10 may be provided with two types of SR HWs. Both types of SR HW may be type-1 with no feature map compression applied and type-2 with feature map compression applied.

In type-1, a 110K slice LUT and a 102K slice register can be used that occupy 45.38% of the total slice LUTs and 21.08% of the total slice registers in the XCKU040 FPGA device.

In Type-2, a 151K slice LUT and a 121K slice register corresponding to 62.6% of the total slice LUT and 24.97% of the total slice register can be used.

In addition, both Type-1 and Type-2 can take full advantage of the 1,920 DSP block in the XCKU040 FPGA device on the KCU105 evaluation board.

By combining feature map compression, Type-2 can reduce on-chip memory usage of about 50% of Type-1 (for example, block RAM on an FPGA). On the other hand, type-2 can further use about 38% slice LUT and about 18% slice register in comparison to type-1 to implement two compressors 2231 and six decompressors 2235.

The image processing apparatus 10 may require a larger number of line memories and gates as compared to the non-CNN based SR method, but it can recover a fairly high quality 4K UHD HR image in real time at 60 fps.

The method 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, and the like, alone or in combination. Program instructions to be recorded on the medium may be those specially designed and constructed for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Magneto-optical media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code 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 the operations of the embodiments, and vice versa.

The software may include a computer program, code, instructions, or a combination of one or more of the foregoing, and may be configured to configure the processing device to operate as desired or to process it collectively or collectively Device can be commanded. The software and / or data may be in the form of any type of machine, component, physical device, virtual equipment, computer storage media, or device , Or may be permanently or temporarily embodied in a transmitted signal wave. The software may be distributed over a networked computer system and stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

Although the embodiments have been described with reference to the drawings, various technical modifications and variations may be applied to those skilled in the art. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (28)

1. A receiver for receiving two-dimensional pixel information;

   At least one first line buffer for outputting the two-dimensional pixel information as a data line on a line basis;

   An arithmetic unit for performing a convolution operation based on the data line data; And

   And at least one second line buffer

   And the image processing apparatus.
2. The method according to claim 1,

   The receiver includes:

   A first pixel information converter for converting the first color data of the two-dimensional pixel information into second color data,

   And the image processing apparatus.
3. The method according to claim 1,

   A second pixel information converter for converting the second color data of the data line on which the operation is performed to first color data,

   The image processing apparatus further comprising:
4. The method according to claim 1,

   The computing unit includes:

   A first convolution operator for generating a feature map by performing a convolution operation based on the data line;

   A feature map processor for storing the feature map in units of at least one line and processing for outputting the feature maps stored in the at least one line unit in a two-dimensional form; And

   A second convolution operator for performing a convolution operation based on the feature map output in a two-

   And the image processing apparatus.
5. 5. The method of claim 4,

   Wherein the first convolution operator comprises:

   (Residual Block) for learning and outputting a residual signal

   Image processing device.
6. 5. The method of claim 4,

   Wherein the first convolution operator comprises:

   At least one 1-D convolution operator for performing 1-D (1-dimensional)

   And the image processing apparatus.
7. 5. The method of claim 4,

   Wherein the first convolution operator comprises:

   A depth-wise convolution operator; And

   A point-wise convolutional arithmetic unit directly connected to the depth-wise convolutional arithmetic unit

   And the image processing apparatus.
8. 5. The method of claim 4,

   Wherein the feature map processor comprises:

   A compressor for compressing the feature map in units of at least one line,

   And the image processing apparatus.
9. 9. The method of claim 8,

   Wherein the feature map processor comprises:

   At least one third line buffer for storing a feature map compressed in units of lines;

   The image processing apparatus further comprising:
10. 10. The method of claim 9,

    Wherein the feature map processor comprises:

    A decompressor for restoring the feature map compressed in the line unit into a two-dimensional feature map,

    The image processing apparatus further comprising:
11. 5. The method of claim 4,

    Wherein the second convolution operator comprises:

    At least one 2-D convolution operator for performing 2-D (2-dimensional)

    And the image processing apparatus.
12. 5. The method of claim 4,

    Wherein the second convolution operator comprises:

    A depth-wise convolution operator; And

    A point-wise convolutional arithmetic unit directly connected to the depth-wise convolutional arithmetic unit

    And the image processing apparatus.
13. 5. The method of claim 4,

    The computing unit includes:

    A quantizer for quantizing the result of at least one convolution operation;

    The image processing apparatus further comprising:
14. 5. The method of claim 4,

    The computing unit includes:

    Weight buffer for storing parameters used for convolution operations

    The image processing apparatus further comprising:
15. Receiving two-dimensional pixel information;

    Outputting the two-dimensional pixel information as at least one data line on a line basis;

    Performing a convolution operation based on the at least one data line; And

    Outputting the data line on which the operation is performed;

    / RTI >
16. 16. The method of claim 15,

    Wherein the receiving comprises:

    Converting the first color data of the two-dimensional pixel information into second color data

    / RTI >
17. 16. The method of claim 15,

    Converting the second color data of the data line on which the operation is performed into first color data

    The image processing method further comprising:
18. 16. The method of claim 15,

    Wherein the performing comprises:

    Generating a feature map by performing a first convolution operation based on the data line;

    Processing the feature map to store the feature map on a line-by-line basis, and output the feature map stored in the at least one line unit in a two-dimensional form; And

    Performing a second convolution operation based on the feature map output in a two-dimensional form

    / RTI >
19. 19. The method of claim 18,

    Wherein the first convolution operation comprises:

    Is performed in a residual block for learning and outputting a residual signal

    Image processing method.
20. 19. The method of claim 18,

    Wherein the generating comprises:

    Performing at least one 1-D (1-dimensional) convolution operation

    / RTI >
21. 19. The method of claim 18,

    Wherein the generating comprises:

    Performing a depth-wise convolution operation; And

    Point-wise convolution operation of the result of the depth-wise convolution operation

    / RTI >
22. 19. The method of claim 18,

    Wherein the processing comprises:

    Compressing the feature map in units of at least one line;

    / RTI >
23. 23. The method of claim 22,

    Wherein the processing comprises:

    Storing the compressed feature map in at least one line unit

    The image processing method further comprising:
24. 24. The method of claim 23,

    Wherein the processing comprises:

    Reconstructing the feature map compressed by the at least one line unit into a two-dimensional feature map

    The image processing apparatus further comprising:
25. 19. The method of claim 18,

    Wherein performing the second convolution operation comprises:

    Performing at least one 2-D (2-dimensional) convolution operation

    / RTI >
26. 19. The method of claim 18,

    Wherein performing the second convolution operation comprises:

    Performing a depth-wise convolution operation; And

    Point-wise convolution operation of the result of the depth-wise convolution operation

    / RTI >
27. 19. The method of claim 18,

    Wherein performing the convolution operation comprises:

    Quantizing the result of at least one convolution operation

    The image processing method further comprising:
28. 19. The method of claim 18,

    Wherein performing the convolution operation comprises:

    Storing the parameters used in the convolution operation

    The image processing method further comprising:

Images