KR102589555B1 - Method for selecting spectral bandwidth of hyperspectral image and spectral bandwidth selection apparatus using the same - Google Patents

Abstract

This application relates to a method of selecting a spectral band of a hyperspectral image and a spectral band setting device using the same, comprising the steps of: receiving a hyperspectral image taken of a fire scene with a hyperspectral camera; Extracting an i-th spectral band and a j-th spectral band as a band pair among the N spectral bands included in the hyperspectral image, calculating a Normalized Difference Index (NDI) for the band pair, and generating a result image; Setting M thresholds and applying each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds; Compare pixels detected as flames in the binarized image with pixels corresponding to actual flames, calculate the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and calculate the FPR of the M binarized images. and generating a Receiver Operation Characteristic (ROC) curve using TPR; Calculating an Area Under the Curve (AUC) value for the ROC curve and setting it as the AUC index for the band pair; And calculating the AUC index for all N You can.

Description

Method for selecting spectral band of hyperspectral image and spectral bandwidth selection apparatus using the same {Method for selecting spectral bandwidth of hyperspectral image and spectral bandwidth selection apparatus using the same}

This application relates to a method for selecting spectral bands of hyperspectral images for real-time fire detection and a spectral band setting device using the same.

Hyperspectral cameras, unlike regular RGB cameras, can generate hyperspectral images containing hundreds of spectral bands in one pixel. Hyperspectral images have three characteristics: many spectral bands, continuous, and narrow wavelength.

In relation to fire detection, unlike existing RGB cameras, hyperspectral images can distinguish different objects even if they are of the same color. Therefore, when detecting a fire using an existing RGB camera, there may be cases where a fire is incorrectly detected due to lighting, sunlight, reflected light, etc. However, when hyperspectral imaging is used, the light source from the screen can be detected by other lighting. It is possible to easily distinguish between.

However, hyperspectral images are large amounts of data with hundreds of bands, and when used as is for fire detection, problems such as time delays during data processing may occur.

This application seeks to provide a spectral band selection method for hyperspectral images that enables real-time processing in a hyperspectral camera and can select a spectral band for data capacity reduction, and a spectral band setting device using the same.

This application seeks to provide a method for selecting a spectral band of a hyperspectral image that can be used for fire detection and a spectral band setting device using the same.

This application seeks to provide a method for selecting a spectral band of a hyperspectral image that can be applied to a line scan type hyperspectral camera and a snapshot type multispectral camera, and a spectral band setting device using the same.

The method of selecting a spectral band of a hyperspectral image according to an embodiment of the present invention relates to a method of selecting a spectral band of a hyperspectral image using a spectral band setting device, and receives a hyperspectral image captured from a fire scene with a hyperspectral camera. steps; Extracting an i-th spectral band and a j-th spectral band as a band pair among the N spectral bands included in the hyperspectral image, calculating a Normalized Difference Index (NDI) for the band pair, and generating a result image; Setting M thresholds and applying each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds; Compare pixels detected as flames in the binarized image with pixels corresponding to actual flames, calculate the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and calculate the FPR of the M binarized images. and generating a Receiver Operation Characteristic (ROC) curve using TPR; Calculating an Area Under the Curve (AUC) value for the ROC curve and setting it as the AUC index for the band pair; And receiving the AUC index for all N You can.

Here, in the step of receiving the hyperspectral image, the hyperspectral image may be captured using a VNIR (Visible and Near Infrared) line-scan hyperspectral camera.

Here, the step of generating the result image is

The NDI is calculated using , and img(i) and img(j) may be the spectral image of the i-th spectral band and the spectral image of the j-th spectral band, respectively.

Here, the step of individually generating the M binarized images sets at least 200 threshold values, and the threshold values may have the same interval from the lowest value to the highest value.

Here, the step of generating the ROC curve is

Calculated using , where TP (True Positive) is the number of pixels corresponding to actual flames among pixels marked as flames in the binarized image, and FN (False Negatives) is the number of pixels corresponding to actual flames among pixels marked as non-flames in the binarized image. FP (False Positive) is the number of pixels corresponding to flames, FP (False Positive) is the number of pixels that do not correspond to actual flames among the pixels marked as flames in the binarized image, and TN (True Negative) is the number of pixels marked as not flames in the binarized image. Among these, the number of pixels may not correspond to actual flames.

Here, in the step of generating the ROC curve, each FPR and TPR for the binarized image are set to coordinate values of the x-axis and y-axis, respectively, and M coordinate points according to the coordinate values are placed on a two-dimensional coordinate plane. By displaying, the ROC curve can be generated.

Here, the step of setting the fire detection band pair is to set the i-th spectral band and the j-th spectral band to coordinate values of the x-axis and y-axis, respectively, and set the AUC index of the band pair corresponding to the coordinate values to color or contrast. By indicating , the AUC matrix can be generated.

Here, in the step of setting the fire detection band pair, the band pair having the maximum AUC index among the N × N AUC indices may be set as the fire detection band pair.

Here, in the step of setting the fire detection band pair, the fire detection band pair is divided into a VNIR (Visible and Near Infrared) region and a NIR (Near Infrared) region, and the band with the AUC index has the maximum value in each region. The pair can be set as the fire detection band pair.

According to an embodiment of the present invention, there may be a computer program stored in a medium combined with hardware to execute the method for selecting a spectral band of a hyperspectral image described above.

Additionally, according to an embodiment of the present invention, there may be a real-time fire detection device that performs fire detection in real time using the fire detection band pair set according to the spectral band selection method of the hyperspectral image described above.

A spectral band setting device according to an embodiment of the present invention includes an image receiving unit that receives a hyperspectral image taken of a fire scene with a hyperspectral camera; Out of the N spectral bands included in the hyperspectral image, the i-th spectral band and the j-th spectral band are extracted as a band pair, and the NDI (Normalized Difference Index) for the band pair is calculated to generate a result image. wealth; a binarized image generator that sets M thresholds and applies each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds; Compare pixels detected as flames in the binarized image with pixels corresponding to actual flames, calculate the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and calculate the FPR of the M binarized images. and a ROC curve generator that generates a Receiver Operation Characteristic (ROC) curve using TPR; An AUC index setting unit that calculates an AUC (Area Under the Curve) value for the ROC curve and sets it as the AUC index for the band pair; And a setting unit that receives the AUC index for all N × N band pairs selectable from the hyperspectral images, compares the AUC index, and sets a fire detection band pair used for real-time fire detection among the band pairs. You can.

Additionally, the means for solving the above problems do not enumerate all the features of the present invention. The various features of the present invention and the resulting advantages and effects can be understood in more detail by referring to the specific embodiments below.

According to a method for selecting a spectral band of a hyperspectral image and a spectral band setting device using the same according to an embodiment of the present invention, it is possible to set the optimal spectral band combination that can be used in a snapshot type hyperspectral camera. In other words, by setting the optimal spectral band combination that can be applied to a snapshot hyperspectral camera, fire detection can be performed in real time and more effective fire detection is possible compared to existing RGB cameras.

However, the effects that can be achieved by the spectral band selection method of hyperspectral images and the spectral band setting device using the same according to embodiments of the present invention are not limited to those mentioned above, and other effects not mentioned are as follows. From the description, it will be clearly understandable to those skilled in the art to which the present invention pertains.

Figure 1 is a schematic diagram showing a fire detection system according to an embodiment of the present invention.
Figures 2 and 3 are block diagrams showing a spectral band setting device according to an embodiment of the present invention.
Figure 3 is a schematic diagram showing target image capturing and target area setting according to an embodiment of the present invention.
Figure 4 is an exemplary diagram showing a fire scene and resulting image according to an embodiment of the present invention.
Figure 5 is an exemplary diagram showing a confusion matrix and ROC curve according to an embodiment of the present invention.
Figure 6 is an exemplary diagram showing an AUC matrix according to an embodiment of the present invention.
Figure 7 is a flowchart showing a method for selecting a spectral band of a hyperspectral image according to an embodiment of the present invention.

Hereinafter, embodiments disclosed in the present specification will be described in detail with reference to the attached drawings. However, identical or similar components will be assigned the same reference numbers regardless of reference numerals, and duplicate descriptions thereof will be omitted. The suffixes “module” and “part” for components used in the following description are given or used interchangeably only for the ease of preparing the specification, and do not have distinct meanings or roles in themselves. That is, the term 'unit' used in the present invention refers to a hardware component such as software, FPGA, or ASIC, and the 'unit' performs certain roles. However, 'wealth' is not limited to software or hardware. The 'part' may be configured to reside on an addressable storage medium and may be configured to run on one or more processors. Therefore, as an example, 'part' refers to components such as software components, object-oriented software components, class components and task components, processes, functions, properties, procedures, Includes subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functionality provided within the components and 'parts' may be combined into a smaller number of components and 'parts' or may be further separated into additional components and 'parts'.

Additionally, in describing the embodiments disclosed in this specification, if it is determined that detailed descriptions of related known technologies may obscure the gist of the embodiments disclosed in this specification, the detailed descriptions will be omitted. In addition, the attached drawings are only for easy understanding of the embodiments disclosed in this specification, and the technical idea disclosed in this specification is not limited by the attached drawings, and all changes included in the spirit and technical scope of the present invention are not limited. , should be understood to include equivalents or substitutes.

Figure 1 is a schematic diagram showing a fire detection system according to an embodiment of the present invention.

Referring to Figure 1, a fire detection system according to an embodiment of the present invention may include a hyperspectral camera 210, a fire detection device 200, and a spectral band setting device 100.

Hereinafter, a fire detection system according to an embodiment of the present invention will be described with reference to FIG. 1.

The hyperspectral camera 210 can capture a surveillance area and generate a hyperspectral image. The surveillance area can be set in various ways, such as buildings, forests, or roads that are at risk of fire, and depending on the embodiment, a tunnel that is at risk of a large-scale casualty accident in the event of a fire may be set as the surveillance area.

Unlike a general Red-Green-Blue (RGB) camera, the hyperspectral camera 210 can acquire spectral information of tens or hundreds of spectral bands for one pixel. In other words, using the hyperspectral camera 210, it is possible to obtain spatial and spectral information at the same time. Here, in the case of hyperspectral images, the pixel value at each pixel along with the two-dimensional position of each pixel can be expressed as a spectral profile, so the spectral information of each pixel can be expressed in a three-dimensional data cube structure.

Depending on the type, hyperspectral cameras can generate hyperspectral images using a line-scan method or a snap shot method. In the case of the line scan method, spectral information for hundreds of spectral bands can be generated by passing one line at a time through the spectrometer, but it takes a long time to generate a hyperspectral image, making it difficult to secure real-time fire detection. On the other hand, in the case of the snapshot method, a spectral image for a set spectral band is generated using a filter, so the number of spectral bands that can be obtained at once is reduced, but it is possible to generate one frame at a time. Therefore, in the case of the snapshot method, it is possible to generate spectroscopic images in real time. Here, the hyperspectral camera 210 can operate in a snapshot manner, and the generated hyperspectral image can be transmitted to the fire detection device 200 in real time.

The fire detection device 200 can detect whether a fire has occurred within the surveillance area using hyperspectral images. In other words, using hyperspectral images, it is possible to easily distinguish between vehicle lights such as LED (Light Emitting Device), HID (High Intensity Discharge Headlamp), and halogen and general street lights, and in the event of a fire, the spectra of the flame differ from that of general lighting. Because different spectra appear, fires can be detected accurately.

Specifically, unlike other light sources, the spectral band of a flame may have a peak value at 770 nm. This is expressed as a potassium-line or K-line, and the fire detection device 200 can detect a fire using the spectroscopic image of the potassium line, which has a peak spectral value of the flame. In other words, since the flame shines brighter than the ambient light in the spectral band of the potassium line, it is possible to easily detect the fire in the image by utilizing this.

However, since the hyperspectral camera 210 can generate spectral images for hundreds of spectral bands, it is necessary to extract the optimal combination of spectral bands to detect fire in real time. In other words, it is necessary to pre-set the spectral band to be used for fire detection, then acquire the spectral image for the corresponding spectral band using a snapshot method, and detect the fire in the spectral image in real time.

The spectral band setting device 100 can calculate the optimal spectral band combination required for fire detection and provide the calculation to the fire detection device 200. The hyperspectral image obtained through the line scan hyperspectral camera 210 includes hundreds of spectral bands, and if fire detection is performed using all of them, time and resources may be wasted in data processing. Additionally, in the case of the line scan method, it takes a long time to acquire data, so it may be more difficult to perform real-time fire detection.

Therefore, in order to use the snapshot type hyperspectral camera 210, it is necessary to set the optimal spectral band for fire detection. Here, the spectral band setting device 100 can select the optimal spectral band for fire detection by evaluating the detection performance of each band pair using the AUC index.

Meanwhile, in Figure 1, the spectral band setting device 100 is shown as being provided separately from the fire detection device 200, but depending on the embodiment, the spectral band setting device 100 is integrated within the fire detection device 200. It is also possible to include In other words, the band setting device 100 can operate as a function of the fire detection device 200.

Hereinafter, a spectral band setting device 100 according to an embodiment of the present invention will be described.

Figure 2 is a block diagram showing a spectral band setting device 100 according to an embodiment of the present invention.

Referring to Figure 2, the spectral band setting device 100 according to an embodiment of the present invention includes an image receiver 110, a result image generator 120, a binarized image generator 130, and an ROC curve generator 140. , may include an AUC index setting unit 150 and a setting unit 160.

The image receiver 110 can receive a hyperspectral image captured at a fire scene using a hyperspectral camera. In other words, in order to set the optimal spectral band for fire detection, a simulated fire scene can be implemented as shown in Figure 4(a) and then photographed with a hyperspectral camera, and the hyperspectral image generated through this can be used with a spectral band setting device. It can be provided as (100). In this case, the spectral band setting device 100 can receive the corresponding hyperspectral image through the image receiver 110. Here, the hyperspectral image received by the image receiver 110 may be generated by a hyperspectral camera using a line-scan method for the VNIR (Visible and Near Infrared) region, and within the hyperspectral image, N Spectral information for spectral bands (N is a natural number) may be included. Depending on the embodiment, N may be at least 500 or more.

The result image generator 120 can extract the i-th spectral band and the j-th spectral band as one band pair among the N spectral bands included in the received hyperspectral image (where i and j are natural numbers). , the resulting image can be generated by calculating the NDI (Normalized Difference Index) for the extracted band pair. Hyperspectral images can be generated in the VNIR region, where the spectral band can be 400 to 1400 nm. Here, the N spectral bands included in the hyperspectral image are the VNIR region divided into N equal parts, and can be sequentially numbered from 1 to N from the lowest spectral band (400 nm) to the highest spectral band (1400 nm). . Therefore, the i-th spectral band and the j-th spectral band can be specified, and the result image generator 120 can select any i and j to extract the corresponding spectral band as a band pair.

Afterwards, the result image generator 120 can extract spectral images corresponding to the extracted i-th spectral band and j-th spectral band, and calculate the corresponding NDI using the formula below.

Here, img(i) and img(j) may be a spectroscopic image of the i-th spectral band and a spectroscopic image of the j-th spectral band, respectively. In this case, the resulting image as shown in Figure 4(b) can be extracted from the fire scene image shown in Figure 4(a). The result image generator 120 may extract a total of N × N selectable band pairs for the N spectral bands and generate each result image.

The binarized image generator 130 may set M thresholds and apply each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds. Since areas with high pixel values in the resulting image correspond to flames, the fire area included in the image can be clearly extracted by converting the resulting image into a binarized image using a threshold value. That is, pixels above the threshold are displayed as 1, and pixels below the threshold are displayed as 0, thereby converting the image into a binarized image.

However, the threshold can be set in various ways, and it may not be desirable for the performance of the selected band pair to be judged differently depending on the threshold setting. Therefore, the binarized image generator 130 can generate binarized images by applying a plurality of threshold values to one result image, and then utilizes all of the multiple binarized images generated for one result image to improve performance. You can compare.

Depending on the embodiment, the binarized image generator 130 may set at least 200 threshold values, and each threshold value may be generated at equal intervals from the lowest value to the highest value. Additionally, the binarized image generator 130 may apply the same threshold values to the resulting images of all N × N band pairs.

The ROC curve generator 140 compares pixels detected as flames in the binarized image with pixels corresponding to the actual flame, calculates the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and M A ROC (Receiver Operation Characteristic) curve can be generated using the FPR and TPR of the binarized images.

Here, the positions of pixels corresponding to the flame in the hyperspectral image received by the image receiver 110 may be specified and known in advance. Accordingly, the ROC curve generator 140 can determine whether pixels detected as flames in each binarized image correspond to pixels corresponding to actual flames.

Specifically, the ROC curve generator 130

You can calculate each FPR and TPR using . At this time, True Positive (TP), False Negatives (FN), False Positive (FP), and True Negative (TN) can be calculated using the confusion matrix of FIG. 5. Specifically, TP may be the number of pixels corresponding to actual flames among pixels marked as flames in the binarized image, and FN may be the number of pixels corresponding to actual flames among pixels marked as non-flames in the binarized image. In addition, FP corresponds to the number of pixels that do not correspond to actual flames among pixels marked as flames in the binarized image, and TN corresponds to the number of pixels that do not correspond to actual flames among pixels marked as non-flames in the binarized image.

When calculating the TPR and FPR for one binarized image in this way, the ROC curve generator 130 sets each FPR and TPR to coordinate values of the x-axis and y-axis, and sets the corresponding coordinate point to two-dimensional coordinates. It can be displayed on a flat surface. Afterwards, TPR and FPR can be calculated in the same way for each binarized image corresponding to the M thresholds, and FPR and TPR can be displayed on a two-dimensional coordinate plane. In this case, M coordinate points can be displayed on a two-dimensional coordinate plane, and an ROC curve can be generated, as shown in Figure 5(b). Here, the larger the number of thresholds, the more accurate the ROC curve can be generated.

The AUC index setting unit 150 can calculate the AUC (Area Under the Curve) value for the ROC curve, and set the calculation result as the AUC index for the corresponding band pair. Here, the AUC value corresponds to the area at the bottom of the ROC curve, and as the AUC value approaches 1, it may mean that the detection performance of the corresponding band pair is higher. Therefore, after calculating the AUC value corresponding to each band pair, this can be set as the AUC index of the corresponding band pair. The AUC index setting unit 150 can calculate the AUC value for all N × N band pairs and set it as the AUC index.

The setting unit 160 can receive the AUC index for all N You can.

Specifically, the setting unit 160 may set the band pair with the maximum AUC index among the N × N AUC indices as the fire detection band pair. That is, the larger the AUC index, the better the detection performance for flames included in the image. Therefore, the setting unit 160 can set the band pair with the maximum AUC index as the fire detection band pair. Depending on the embodiment, a band pair with an AUC index greater than a set value may be extracted and set as a candidate group for fire detection band pairs, and it is also possible to set a band pair selected by the user among the candidate groups as a fire detection band pair.

In addition, the setting unit 160 distinguishes between the selected band pairs that are included in the VNIR (Visible and Near Infrared) region or the NIR (Near Infrared) region, and detects fire the band pair with the highest AUC index in each region. It is also possible to set it as a band pair. In this case, in the VNIR region, the ith spectral band is 610~930nm, the jth spectral band is between 430~517nm, and the fire detection band pair can be set. In the NIR region, the ith spectral band is 770~920nm, j. As the second spectral band, the spectral band between 940 and 1000 nm can be set as a fire detection band pair.

In addition, depending on the embodiment, the setting unit 160 may also generate an AUC matrix as shown in FIG. 6. That is, the i-th spectral band and the j-th spectral band can be set to coordinate values of the x-axis and y-axis, respectively, and the AUC index of the band pair corresponding to the coordinate value is displayed in color or brightness, thereby creating an AUC matrix. In this case, the user can clearly visually recognize the AUC index for each band pair by utilizing the AUC matrix. At this time, when the user selects a band pair through the AUC matrix, it is also possible for the setting unit 160 to set the corresponding band pair as a fire detection band pair. Here, referring to Figure 6, if the i-th spectral band corresponds to 770 nm, which is the potassium line, it can be confirmed that the j-th spectral band has a high AUC index value no matter which one is selected.

The setting unit 160 may provide a set fire detection band pair to the fire detection device 200, and the fire detection device 200 may use a snapshot-type hyperspectral camera 210 according to the received fire detection band pair. You can set the spectral band of the acquired spectral image. Through this, the fire detection device 200 is capable of performing fire detection in real time using hyperspectral images.

Meanwhile, as shown in Figure 3, the band setting device 100 according to an embodiment of the present invention may include physical components such as a processor 10 and a memory 40, and the memory 40 It may include one or more modules configured to be executed by the processor 10. Specifically, one or more modules may include an image reception module, a result image generation module, a binarized image generation module, an ROC curve generation module, an AUC index setting module, and a setting module.

The processor 10 may execute various software programs and instruction sets stored in the memory 40 to perform various functions and process data. The peripheral interface unit 30 can connect input/output peripheral devices of the computer device to the processor 10 and memory 40, and the memory controller 20 allows the processor 10 or components of the computer device to connect the memory 40. When accessing, a function to control memory access can be performed. Depending on the embodiment, the processor 10, memory controller 20, and peripheral interface unit 30 may be implemented on a single chip or may be implemented as separate chips.

Memory 40 may include high-speed random access memory, one or more magnetic disk storage devices, non-volatile memory such as flash memory devices, etc. Additionally, the memory 40 may further include a storage device located away from the processor 10 or a network attached storage device accessed through a communication network such as the Internet.

As shown in Figure 3, the band setting device 100 includes an operating system in the memory 40, an image reception module corresponding to an application program, a result image generation module, a binarization image generation module, an ROC curve generation module, and an AUC. It may include index setting module and setting module. Here, each module is a set of instructions for performing the above-described functions and can be stored in the memory 40. Accordingly, in the terminal device 100 according to an embodiment of the present invention, the processor 10 can access the memory 40 and execute instructions corresponding to each module.

However, the image receiving module, result image generation module, binarization image generation module, ROC curve generation module, AUC index setting module and setting module are the above-mentioned image receiving unit, result image generation unit, binarization image generation unit, ROC curve generation unit, and AUC Since the index setting unit and setting unit correspond to each other, detailed description is omitted here.

Figure 7 is a flowchart showing a method for selecting a spectral band of a hyperspectral image according to an embodiment of the present invention.

Referring to Figure 7, the method for selecting a spectral band of a hyperspectral image according to an embodiment of the present invention includes a hyperspectral image receiving step (S10), a resulting image generating step (S20), a binarized image generating step (S30), and an ROC curve. It may include a creation step (S40), an AUC index setting step (S50), and a fire detection band pair setting step (S60).

Hereinafter, a method for selecting a spectral band of a hyperspectral image according to an embodiment of the present invention will be described with reference to FIG. 7.

In the hyperspectral image reception step (S10), a hyperspectral image captured by a hyperspectral camera of a fire scene can be received. In other words, after implementing a simulated fire scene, it can be filmed with a hyperspectral camera, and the hyperspectral image generated through this can be received. Depending on the embodiment, the hyperspectral camera may generate hyperspectral images using a line-scan method for the VNIR (Visible and Near Infrared) region, and there are N hyperspectral images (N is a natural number). ) may include spectral information about the spectral band. Here, N may be at least 500 or more.

In the resulting image generation step (S20), the i-th spectral band and the j-th spectral band are extracted as a band pair among the N spectral bands included in the hyperspectral image, and the NDI (Normalized Difference Index) for the band pair is calculated. The resulting image can be generated. Specifically, in the result image generation step (S20), spectral images corresponding to the extracted i-th spectral band and j-th spectral band can be extracted, and the corresponding NDI can be calculated using the formula below.

Here, img(i) and img(j) may be a spectroscopic image of the i-th spectral band and a spectroscopic image of the j-th spectral band, respectively. In the result image generation step (S20), a total of N × N selectable band pairs for N spectral bands can be extracted to generate each result image.

In the binarized image generation step (S30), M threshold values are set and each threshold value is applied to the resulting image, so that M binarized images corresponding to the threshold values can be individually generated. Since the area with high pixel value in the resulting image corresponds to the area corresponding to the flame, by generating a binarized image using the threshold value, the fire area included in the image can be clearly extracted. That is, pixels above the threshold are displayed as 1, and pixels below the threshold are displayed as 0, thereby converting the image into a binarized image.

However, the threshold can be set in various ways, and it may not be desirable for the performance of the selected band pair to be judged differently depending on the threshold setting. Therefore, in the binarization image generation step (S30), each binarization image can be generated by applying a plurality of threshold values to one result image, and then performance is improved by utilizing all of the plurality of binarization images generated for one result image. You can compare. Depending on the embodiment, at least 200 threshold values may be set, and each threshold value may be set at the same interval from the lowest value to the highest value.

In the ROC curve generation step (S40), the pixels detected as flames in the binarized image are compared with the pixels corresponding to the actual flame, and the FPR (False Positive Rate) and TPR (True Positive Rate) of the binarized image can be calculated. . Afterwards, a ROC (Receiver Operation Characteristic) curve can be generated using the FPR and TPR of the M binarized images.

Here, the positions of pixels corresponding to the flame in the received hyperspectral image may be specified and known in advance. Therefore, in the ROC curve generation step (S40), it is possible to determine whether pixels detected as flames in each binarized image correspond to pixels corresponding to actual flames.

Specifically, You can calculate each FPR and TPR using . At this time, True Positive (TP), False Negatives (FN), False Positive (FP), and True Negative (TN) can be calculated using the confusion matrix of FIG. 5.

After calculating the TPR and FPR for one binarized image in this way, each FPR and TPR can be set as coordinate values of the x-axis and y-axis, and the corresponding coordinate point can be displayed on a two-dimensional coordinate plane. Additionally, the TPR and FPR can be calculated in the same way for each binarized image corresponding to the M thresholds, and the M FPRs and TPRs can be displayed on a two-dimensional coordinate plane. In this case, M coordinate points can be displayed on a two-dimensional coordinate plane.

In the AUC index setting step (S50), the AUC (Area Under the Curve) value for the ROC curve can be calculated and set as the AUC index for the band pair. Here, the AUC value corresponds to the area at the bottom of the ROC curve, and as the AUC value approaches 1, it may mean that the detection performance of the corresponding band pair is higher. Therefore, the AUC value corresponding to each band pair can be calculated and then set as the AUC index of the corresponding band pair. In the AUC index setting step (S50), the AUC value can be calculated for all N × N band pairs and set as the AUC index.

In the fire detection band pair setting step (S60), the AUC index is received for all N You can set it.

Specifically, among N × N AUC indices, the band pair with the highest AUC index can be set as the fire detection band pair. In other words, the larger the AUC index, the better the detection performance for flames included in the image. Therefore, the band pair with the maximum AUC index can be set as the fire detection band pair. Depending on the embodiment, a band pair with an AUC index greater than a set value may be extracted and set as a candidate group for fire detection band pairs, and it is also possible to set a band pair selected by the user among the candidate groups as a fire detection band pair.

In addition, in the fire detection band pair setting step (S60), among the selected band pairs, those included in the VNIR (Visible and Near Infrared) area or the NIR (Near Infrared) area are distinguished, and the AUC index in each area has the maximum value. It is also possible to set the band pair as a fire detection band pair.

In addition, depending on the embodiment, it is also possible to generate an AUC matrix as shown in FIG. 6 in the fire detection band pair setting step (S60). That is, the i-th spectral band and the j-th spectral band can be set to coordinate values of the x-axis and y-axis, respectively, and the AUC index of the band pair corresponding to the coordinate value is displayed in color or brightness, thereby creating an AUC matrix. In this case, the user can clearly visually recognize the AUC index for each band pair by using the AUC matrix, and when the user selects a band pair through the AUC matrix, the corresponding band pair is set as a fire detection band pair. You can do it.

The above-described present invention can be implemented as computer-readable code on a program-recorded medium. A computer-readable medium may continuously store a computer-executable program or temporarily store it for execution or download. In addition, the medium may be a variety of recording or storage means in the form of a single or several pieces of hardware combined. It is not limited to a medium directly connected to a computer system and may be distributed over a network. Examples of media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, And there may be something configured to store program instructions, including ROM, RAM, flash memory, etc. Additionally, examples of other media include recording or storage media managed by app stores that distribute applications, sites or servers that supply or distribute various other software, etc. Accordingly, the above detailed description should not be construed as restrictive in all respects and should be considered illustrative. The scope of the present invention should be determined by reasonable interpretation of the appended claims, and all changes within the equivalent scope of the present invention are included in the scope of the present invention.

The present invention is not limited to the above-described embodiments and attached drawings. For those skilled in the art to which the present invention pertains, it will be clear that components according to the present invention can be replaced, modified, and changed without departing from the technical spirit of the present invention.

100: Spectral band setting device 110: Image receiving unit
120: Result image generation unit 130: Binarization image generation unit
140: ROC curve generation unit 150: AUC index setting unit
160: Setting unit 200: Fire detection device
210: Hyperspectral camera
S10: Hyperspectral image reception step S20: Resulting image generation step
S30: Binarization image generation step S40: ROC curve generation step
S50: AUC index setting step S60: Fire detection band pair setting step

Claims (12)

In the method of selecting a spectral band of a hyperspectral image using a spectral band setting device,
Receiving a hyperspectral image captured at a fire scene with a hyperspectral camera;
Extracting an i-th spectral band and a j-th spectral band as a band pair among the N spectral bands included in the hyperspectral image, calculating a Normalized Difference Index (NDI) for the band pair, and generating a result image;
Setting M thresholds and applying each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds;
Compare pixels detected as flames in the binarized image with pixels corresponding to actual flames, calculate the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and calculate the FPR of the M binarized images. and generating a Receiver Operation Characteristic (ROC) curve using TPR;
Calculating an Area Under the Curve (AUC) value for the ROC curve and setting it as the AUC index for the band pair; and
Comprising the step of receiving the AUC index for a total of N Spectral band selection method for hyperspectral images.
The method of claim 1, wherein receiving the hyperspectral image includes
A spectral band selection method for a hyperspectral image, characterized in that the hyperspectral image is captured using a VNIR (Visible and Near Infrared) line-scan hyperspectral camera.
The method of claim 1, wherein the step of generating the result image is

Calculate the NDI using, and img(i) and img(j) are the spectral image of the i-th spectral band and the spectral image of the j-th spectral band, respectively. Spectral band selection method of hyperspectral image. .
The method of claim 1, wherein the step of individually generating the M binarized images includes
A spectral band selection method for a hyperspectral image, wherein at least 200 threshold values are set, and the threshold values are spaced equally from the lowest value to the highest value.
The method of claim 1, wherein generating the ROC curve

Calculated using , where TP (True Positive) is the number of pixels corresponding to actual flames among pixels marked as flames in the binarized image, and FN (False Negatives) is the number of pixels corresponding to actual flames among pixels marked as non-flames in the binarized image. This is the number of pixels corresponding to the flame,
FP (False Positive) is the number of pixels that do not correspond to an actual flame among the pixels marked as flames in the binarized image, and TN (True Negative) is the number of pixels that do not correspond to an actual flame among the pixels that are marked as not a flame in the binarized image. A spectral band selection method for a hyperspectral image, characterized in that:
The method of claim 1, wherein generating the ROC curve
Setting each FPR and TPR for the binarized image to coordinate values of the x-axis and y-axis, respectively, and displaying M coordinate points according to the coordinate values on a two-dimensional coordinate plane to generate the ROC curve. Characterized by spectral band selection method for hyperspectral images.
The method of claim 1, wherein the step of setting the fire detection band pair is
The i-th spectral band and the j-th spectral band are set to coordinate values of the x-axis and y-axis, respectively, and the AUC index of the band pair corresponding to the coordinate value is displayed in color or contrast to generate an AUC matrix. Method for selecting spectral bands in hyperspectral images.
The method of claim 1, wherein the step of setting the fire detection band pair is
A spectral band selection method for a hyperspectral image, characterized in that, among the N × N AUC indices, the band pair having the maximum AUC index is set as the fire detection band pair.
The method of claim 1, wherein the step of setting the fire detection band pair is
The fire detection band pair is divided into a VNIR (Visible and Near Infrared) region and a NIR (Near Infrared) region, and the band pair with the maximum AUC index in each region is set as the fire detection band pair. Method for selecting spectral bands of transparticle images.
A computer program combined with hardware and stored on a medium to execute the method of selecting a spectral band of a hyperspectral image according to any one of claims 1 to 9.
A real-time fire detection device that performs fire detection in real time using the fire detection band pair set according to the spectral band selection method of the hyperspectral image of any one of claims 1 to 9.
An image receiving unit that receives hyperspectral images taken of a fire scene with a hyperspectral camera;
Out of the N spectral bands included in the hyperspectral image, the i-th spectral band and the j-th spectral band are extracted as a band pair, and the NDI (Normalized Difference Index) for the band pair is calculated to generate a result image. wealth;
a binarized image generator that sets M thresholds and applies each threshold to the resulting image to individually generate M binarized images corresponding to the thresholds;
Compare pixels detected as flames in the binarized image with pixels corresponding to actual flames, calculate the false positive rate (FPR) and true positive rate (TPR) of the binarized image, and calculate the FPR of the M binarized images. and a ROC curve generator that generates a Receiver Operation Characteristic (ROC) curve using TPR;
An AUC index setting unit that calculates an AUC (Area Under the Curve) value for the ROC curve and sets it as the AUC index for the band pair; and
Spectroscopy comprising a setting unit that receives the AUC index for all N Band setting device.

Images