Machine Learning Device

Abstract

The machine learning device (100) includes a resolution reduction processing unit (103) that generates a low-resolution interference map (104) and a low-resolution interference label (105) based on an interference map (101) and an interference label (102), a resolution enhancement processing unit (106) that generates a high-resolution interference map (107) based on the low-resolution interference map (104) and the low-resolution interference label (105) generated by the resolution reduction processing unit (103), and a loss calculation unit (108) that calculates a loss based on the interference map (101), the interference label (102), and the high-resolution interference map (107) generated by the resolution enhancement processing unit (106). The loss calculation unit (108) calculates the loss using a loss function that adds or multiplies weights that change corresponding to pixel values, and trains the resolution enhancement processing unit (106) to minimize the loss.

Description

[Technical Field]

[0001] The present disclosure relates to a machine learning device and a high-resolution interference map generator used in single-image super-resolution techniques. [Background technology]

[0002] A single-image super-resolution technique that has attracted attention in recent years and various methods have been proposed is a learning-based method using machine learning or deep learning. Non-Patent Document 1 proposes a method of generating a single-image super-resolution machine learning model by preparing a high-resolution image, converting the high-resolution image into a low-resolution image, and then training a convolutional neural network (CNN).

[0003] Patent Document 1 proposes a method for improving the accuracy of frequency control by super-resolution of an interference map that indicates power for each time frequency used in a wireless communication system, and combining this with an interference source classification method. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7214059 [Non-patent literature]

[0005] [Non-Patent Document 1] Chao Dong et al., “Image Super-Resolution Using Deep Convolutional Networks”, arXiv:1501.00092v3 [cs.CV], published July 31, 2015 Summary of the Invention [Problem to be solved by the invention]

[0006] Conventional learning-based single-image super-resolution techniques define an error function that accurately restores all pixels and trains to minimize the error.When applying super-resolution to interference maps, it is most important to determine where high-power interference waves exist.Therefore, a machine learning model is needed that does not precisely reproduce low-power noise regions, but more accurately reproduces the interference areas.

[0007] The present disclosure has been made in consideration of the above, and aims to provide a machine learning device that can improve the accuracy of estimating an interference area in an interference map. [Means for solving the problem]

[0008] To solve the above-mentioned problems and achieve the object, the machine learning device according to the present disclosure includes a resolution reduction processing unit that generates a low-resolution interference map and a low-resolution interference label based on an interference map and an interference label, a resolution enhancement processing unit that generates a high-resolution interference map based on the low-resolution interference map and the low-resolution interference label generated by the resolution reduction processing unit, and a loss calculation unit that calculates a loss based on the interference map, the interference label, and the high-resolution interference map generated by the resolution enhancement processing unit. The loss calculation unit calculates the loss using a loss function that adds or multiplies weights that change corresponding to pixel values, and trains the resolution enhancement processing unit to minimize the loss. [Effects of the Invention]

[0009] The machine learning device according to the present disclosure advantageously achieves the effect of improving the accuracy of estimating an interference area in an interference map. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram showing a configuration of a machine learning device according to a first embodiment. [Figure 2] FIG. 10 is a diagram showing the configuration of a machine learning device according to a second embodiment. [Figure 3] FIG. 10 is a diagram showing the configuration of a high-resolution interference map generating device according to a third embodiment. [Figure 4]FIG. 1 is a diagram showing a processor in a case where some or all of the functions of a resolution reduction processor, a resolution improvement processor, and a loss calculation unit included in a machine learning device according to a first embodiment are implemented by a processor. [Figure 5] FIG. 1 is a diagram showing a processing circuit in the case where some or all of the functions of a resolution reduction processor, a resolution improvement processor, and a loss calculation unit included in a machine learning device according to a first embodiment are realized by the processing circuit. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, a machine learning device and a high-resolution interference map generating device according to an embodiment will be described in detail with reference to the accompanying drawings.

[0012] Embodiment 1 FIG. 1 is a diagram showing a configuration of a machine learning device 100 according to a first embodiment. The machine learning device 100 is a device used in single-image super-resolution technology. The machine learning device 100 has a resolution reduction processor 103 that receives an interference map 101 and an interference label 102 as input and generates a low-resolution interference map 104 and a low-resolution interference label 105 based on the interference map 101 and the interference label 102. The machine learning device 100 further has a resolution enhancement processor 106 that generates a high-resolution interference map 107 based on the low-resolution interference map 104 and the low-resolution interference label 105 generated by the resolution reduction processor 103. The machine learning device 100 further has a loss calculation unit 108 that calculates a loss based on the interference map 101, the interference label 102, and the high-resolution interference map 107 generated by the resolution enhancement processor 106.

[0013] The resolution reduction processor 103 acquires the interference map 101 and the interference label 102 and reduces the number of pixels in accordance with the resolution enhancement factor of the resolution enhancement processor 106. That is, the resolution reduction processor 103 executes a resolution reduction method. Examples of the resolution reduction method include averaging, filtering, or maximum value extraction within a specific region. However, the resolution reduction method is not limited to these. The resolution reduction processor 103 generates a low-resolution interference map 104 and a low-resolution interference label 105. The generated low-resolution interference label 105 is assumed to be adjusted so that the value of each pixel is an integer, but may be a floating-point number depending on the learning algorithm of the resolution enhancement processor 106.

[0014] The high-resolution processing unit 106 is configured with a learning-based machine learning model, and receives the low-resolution interference map 104 and the low-resolution interference labels 105 as input to generate the high-resolution interference map 107. Examples of the machine learning model include a Super Resolution Deep Neural Network (SRDNN), a Super Resolution Convolutional Neural Network (SRCNN), a Super Resolution Generative Adversarial Network (SRGAN), and a Super Resolution Transformer (SRT). However, the machine learning model is not limited to these. When the low-resolution interference map 104 and the low-resolution interference labels 105 are input to the high-resolution processing unit 106, multi-input multimodal learning may be performed in which the low-resolution interference map 104 and the low-resolution interference labels 105 are combined and input in the channel direction. However, the input of the low-resolution interference map 104 and the low-resolution interference labels 105 to the high-resolution processing unit 106 is not limited to this.

[0015] The loss calculation unit 108 calculates the loss using a loss function L that multiplies a weight that changes according to the pixel value based on the interference map 101, the interference label 102, and the high-resolution interference map 107, and trains the high-resolution processing unit 106 to minimize the loss. It is assumed that the loss increases as the difference in pixel value increases. For example, the loss function L may be a function expressed by the following formula (1). However, the loss function L is not limited to the function expressed by formula (1). The loss function L may also be a function that adds a weight that changes according to the pixel value. In other words, the loss calculation unit 108 may add weights instead of multiplying them when calculating the loss.

[0016]

number

[0017] In equation (1), H represents the vertical width of the interference map 101, W represents the horizontal width of the interference map 101, and y ij indicates the pixel value at position (i,j) of the interference map 101, and Y ij indicates the pixel value at position (i, j) of the high-resolution interference map 107, and t ij indicates the label value at the position (i, j) of the low-resolution interference label 105. When the loss calculation unit 108 uses the loss function L expressed by Equation (1), it also uses the low-resolution interference label 105 generated by the resolution reduction processing unit 103. μ (y ij ,Y ij ,t ij ) is y ij , Y ij and t ij The function that returns the weight corresponding to w μ (y ij ,Y ij ,t ij ) can be expressed by the following formula (2). However, w μ (y ij ,Y ij ,t ij ) is not limited to the function expressed by equation (2).

[0018]

number

[0019] In equation (2), μ(t ij ) is the label t ij is an arbitrary weighting factor corresponding to

[0020] The loss function L may be a function expressed by the following formula (3). The loss function L expressed by formula (3) is a function whose content changes depending on the label. The loss function L expressed by formula (3) is a function that allows for an error in an area without interference.

[0021]

number

[0022] In equation (3), I is an indicator function that is 1 if the condition is met and 0 if the condition is not met. c is an arbitrary constant. Here, hinge loss and MSE (Mean Squared Error) are used as losses corresponding to the labels, but MAE (Mean Absolute Error) or KL (Kullback-Leibler) Divergence, etc., can also be used.

[0023] As described above, the loss calculation unit 108 calculates the loss using the loss function L that multiplies or adds weights that change according to pixel values, and trains the high resolution processing unit 106 so as to minimize the loss. That is, the machine learning device 100 according to the first embodiment can cause the high resolution processing unit 106 to learn so as to improve the accuracy of estimating regions that indicate interference present in the interference map. Because the high resolution processing unit 106 learns, the machine learning device 100 can improve the accuracy of estimating interference regions in the interference map.

[0024] Embodiment 2 In the first embodiment, the machine learning device 100 that calculates the loss using interference labels is described. In the second embodiment, a machine learning device that is used in an environment where no labels exist in the training data is described.

[0025] 2 is a diagram showing the configuration of a machine learning device 200 according to embodiment 2. In embodiment 2, differences from the machine learning device 100 according to embodiment 1 will be mainly described, and detailed descriptions of the same parts as those of the machine learning device 100 will be omitted.

[0026] The machine learning device 200 includes a label generation unit 202 that generates an estimated interference label 203 based on an interference map 201, and a resolution reduction processing unit 204 that generates a low-resolution interference map 205 and a low-resolution interference label 206 based on the interference map 201 and the estimated interference label 203 generated by the label generation unit 202. The machine learning device 200 further includes a resolution enhancement processing unit 207 that generates a high-resolution interference map 208 based on the low-resolution interference map 205 and the low-resolution interference label 206 generated by the resolution reduction processing unit 204. The machine learning device 200 further includes a loss calculation unit 209 that calculates a loss based on the interference map 201, the estimated interference label 203 generated by the label generation unit 202, and the high-resolution interference map 208 generated by the resolution enhancement processing unit 207. The resolution reduction processing unit 204 has the same function as the resolution reduction processing unit 103 of the first embodiment, the resolution reduction processing unit 207 has the same function as the resolution reduction processing unit 106 of the first embodiment, and the loss calculation unit 209 has the same function as the loss calculation unit 108 of the first embodiment.

[0027] The label generation unit 202 receives the interference map 201 as an input, estimates labels based on the interference map 201, and generates and outputs estimated interference labels 203. For example, the label generation unit 202 generates the estimated interference labels 203 by indicating that there is interference if the target pixel value is equal to or greater than a threshold, and indicating that there is no interference if the target pixel value is less than the threshold, or by generating the estimated interference labels 203 based on segmentation. Examples of segmentation include semantic segmentation, instance segmentation, and panoptic segmentation. However, the segmentation is not limited to these. Furthermore, the label generation unit 202 may perform a filtering process on the interference map 201 in advance in order to reduce the complexity of the interference region.

[0028] The resolution reduction processor 204 receives the interference map 201 and the estimated interference label 203 as input, and outputs a low-resolution interference map 205 and a low-resolution interference label 206 .

[0029] The high resolution processor 207 receives the low resolution interference map 205 and the low resolution interference label 206 as input, and outputs a high resolution interference map 208 .

[0030] The loss calculation unit 209 calculates the loss based on the interference map 201 , the estimated interference label 203 , and the high-resolution interference map 208 .

[0031] As described above, even when no label corresponding to the interference map 201 exists, the machine learning device 200 according to the second embodiment can internally generate an estimated interference label 203 and input it to the high resolution processing unit 207. As a result, similar to the machine learning device 100, the machine learning device 200 can improve the accuracy of estimating an interference region in the interference map.

[0032] Embodiment 3 In the third embodiment, a system configuration for implementing the super-resolution machine learning model trained in the first or second embodiment will be described.

[0033] Fig. 3 is a diagram showing the configuration of a high-resolution interference map generation apparatus 300 according to embodiment 3. The high-resolution interference map generation apparatus 300 includes a frequency conversion unit 301 that converts the frequency of a received signal in the RF (Radio Frequency) band into a baseband, and an analog-to-digital converter 302 that converts the analog signal into a digital signal by sampling or the like on the baseband analog signal obtained by the frequency conversion unit 301. In Fig. 3, the analog-to-digital converter 302 is denoted as ADC302.

[0034] The high-resolution interference map generation device 300 further includes a symbol rate conversion unit 303 that down-samples the digital signal obtained by the analog-to-digital converter 302 and converts it to a desired symbol rate, and an interference signal extraction unit 304 that acquires interference information in the form of a received signal strength indicator (RSSI) or a complex number from the digital signal obtained by the symbol rate conversion unit 303. The high-resolution interference map generation device 300 further includes an interference information storage memory 305 that stores the interference information acquired by the interference signal extraction unit 304. The interference information storage memory 305 is realized by, for example, a semiconductor memory. The high-resolution interference map generation device 300 further includes an interference map generation unit 306 that generates an interference map based on the interference information stored in the interference information storage memory 305.

[0035] The high-resolution interference map generating device 300 further includes a label generating unit 307 that generates interference labels from the interference map generated by the interference map generating unit 306, and a high-resolution processing unit 308 that creates a high-resolution interference map in which the interference area is reproduced with high accuracy based on the interference map generated by the interference map generating unit 306 and the interference labels generated by the label generating unit 307.

[0036] As described above, the high resolution interference map generation apparatus 300 according to the third embodiment creates an interference map based on interference information extracted from received signals, creates interference labels, and then performs a high resolution process. Therefore, the high resolution interference map generation apparatus 300 can configure a system that can estimate interference regions on an interference map with high accuracy.

[0037] 4 is a diagram illustrating a processor 91 in a case where some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 included in the machine learning device 100 according to the first embodiment are realized by the processor 91. In other words, some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 may be realized by the processor 91 that executes a program stored in a memory 92. The processor 91 is a CPU (Central Processing Unit), a processing system, an arithmetic system, a microprocessor, or a DSP (Digital Signal Processor). The memory 92 is also shown in FIG.

[0038] When some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 are realized by the processor 91, the functions are realized by the processor 91 together with software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 92. The processor 91 realizes some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 by reading and executing the program stored in the memory 92.

[0039] When some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 are realized by the processor 91, the machine learning device 100 has a memory 92 for storing a program that results in the execution of some or all of the steps executed by the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108. It can also be said that the program stored in the memory 92 causes a computer to execute some or all of the procedures or methods executed by the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108.

[0040] The memory 92 may be, for example, a non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (registered trademark) (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disk).

[0041] 5 is a diagram illustrating the processing circuit 93 in a case where some or all of the functions of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 included in the machine learning device 100 according to embodiment 1 are realized by the processing circuit 93. In other words, some or all of the resolution reduction processing unit 103, the resolution enhancement processing unit 106, and the loss calculation unit 108 may be realized by the processing circuit 93.

[0042] The processing circuitry 93 is dedicated hardware, and may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0043] Some of the functions of the resolution reduction processing unit 103, the resolution improvement processing unit 106, and the loss calculation unit 108 may be realized by dedicated hardware that is separate from the hardware that realizes the remaining functions of the resolution reduction processing unit 103, the resolution improvement processing unit 106, and the loss calculation unit 108.

[0044] Some of the functions of the resolution reduction processor 103, the resolution enhancement processor 106, and the loss calculation unit 108 may be implemented by software or firmware, and the remaining functions may be implemented by dedicated hardware. In this way, the functions of the resolution reduction processor 103, the resolution enhancement processor 106, and the loss calculation unit 108 can be implemented by hardware, software, firmware, or a combination of these.

[0045] Some or all of the functions of the label generation unit 202, the resolution reduction processing unit 204, the resolution improvement processing unit 207, and the loss calculation unit 209 included in the machine learning device 200 according to the second embodiment may be implemented by a processor or a processing circuit. The processor is equivalent to the processor 91. The processing circuit is equivalent to the processing circuit 93.

[0046] Some or all of the functions of the analog-to-digital converter 302, the symbol rate conversion unit 303, the interference signal extraction unit 304, the interference map generation unit 306, the label generation unit 307, and the high resolution processing unit 308 included in the high resolution interference map generation apparatus 300 according to the third embodiment may be realized by a processor or a processing circuit. The processor is a processor equivalent to the processor 91. The processing circuit is a processing circuit equivalent to the processing circuit 93.

[0047] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other, or part of the configuration may be omitted or modified within the scope of the gist of the invention. [Explanation of symbols]

[0048] 91 processor, 92 memory, 93 processing circuit, 100, 200 machine learning device, 101, 201 interference map, 102 interference label, 103, 204 low-resolution processing unit, 104, 205 low-resolution interference map, 105, 206 low-resolution interference label, 106, 207, 308 high-resolution processing unit, 107, 208 high-resolution interference map, 108, 209 loss calculation unit, 202, 307 label generation unit, 203 estimated interference label, 300 high-resolution interference map generation device, 301 frequency conversion unit, 302 analog-to-digital converter, 303 symbol rate conversion unit, 304 interference signal extraction unit, 305 memory for storing interference information, 306 interference map generation unit.

Claims

1. a resolution reduction processing unit that generates a low-resolution interference map and a low-resolution interference label based on the interference map and the interference label; a resolution enhancement processor that generates a high-resolution interference map based on the low-resolution interference map generated by the resolution reduction processor and the low-resolution interference label; a loss calculation unit that calculates a loss based on the interference map, the interference label, and the high-resolution interference map generated by the high-resolution processing unit, The loss calculation unit calculates the loss using a loss function that adds or multiplies weights that change according to pixel values, and trains the high-resolution processing unit to minimize the loss. A machine learning device characterized by:

2. a label generation unit that generates estimated interference labels based on the interference map; a resolution reduction processing unit that generates a low-resolution interference map and a low-resolution interference label based on the interference map and the estimated interference label generated by the label generation unit; a resolution enhancement processor that generates a high-resolution interference map based on the low-resolution interference map generated by the resolution reduction processor and the low-resolution interference label; a loss calculation unit that calculates a loss based on the interference map, the estimated interference label generated by the label generation unit, and the high-resolution interference map generated by the high-resolution processing unit, the label generator generates the estimated interference labels by subtracting a threshold from pixel values of interest or based on segmentation; The loss calculation unit calculates the loss using a loss function that adds or multiplies weights that change according to pixel values, and trains the high-resolution processing unit to minimize the loss. A machine learning device characterized by:

3. The loss function is a function expressed by the following formula (1): [Equation 1] In the formula (1), H represents the vertical width of the interference map, W represents the horizontal width of the interference map, and y ｉｊ denotes the pixel value at position (i, j) of the interference map, and Y ｉｊ denotes the pixel value at position (i, j) of the high-resolution interference map, and t ｉｊ indicates the label value at the position (i, j) of the low-resolution interference label, and w μ (y ｉｊ , Y ｉｊ , t ｉｊ ) is y ｉｊ , Y ｉｊ and t ｉｊ Here we show a function that returns the weight corresponding to 3. The machine learning device according to claim 1 or 2.

4. The loss function is a function expressed by the following formula (2): [Equation 2] In the formula (2), H represents the vertical width of the interference map, W represents the horizontal width of the interference map, I represents an indicator function, and t ｉｊ indicates the label value at position (i, j) of the low-resolution interference label, and y ｉｊ denotes the pixel value at position (i, j) of the interference map, and Y ｉｊ denotes the pixel value at position (i, j) of the high-resolution interference map, and c is an arbitrary constant.

3. The machine learning device according to claim 1 or 2.

Images