JP2018197940A - Identification method of captured worms - Google Patents

Abstract

To provide an identification method and an identification system of captured worms that are capable of performing an identification work of worms automatically and rapidly and with a relatively low cot and are capable of obtaining an identification result having objectivity and consistency.SOLUTION: A method includes an image reading step, a worm candidate area acquisition step, a background region removal step, a worm area extraction step, a characteristic extraction step for acquiring Merkmal data of a morphological characteristic from the worm area, and an identification step for identifying the worms by verification of the Merkmal data to the standard Merkmal data. The worm candidate area acquisition step includes a step of converting the RGB value to the HSV color system to remove a white line area existing in a sheet image that becomes an obstacle for extraction of the worm candidate area and a step of performing extraction of the white line area using the Hough transformation after extraction of a candidate of the white line area using a chroma and a brightness.SELECTED DRAWING: Figure 3

Description

  The present invention relates to an identification method and an identification system for captured insects, and more specifically, it is necessary to prevent as much as possible the entry of insects and other small insects such as food factories and pharmaceutical factories. The present invention relates to an identification method and an identification system for captured insects in order to identify captured insects at a place and obtain materials to be used for subsequent insect control measures.

  In food factories and pharmaceutical factories, it is required to prevent as much as possible the entry of insects and other small insects, both harmful and harmless. For this purpose, various types of insect trapping devices including an attracting light for attracting insects and an adhesive means for capturing the gathered insects are deployed at important points in the factory building.

  By the way, there are many kinds of reptiles, and there are thousands of kinds even if known, and the wavelength range that is strongly attracted for each reptile differs. Therefore, in order to efficiently capture the insects to be captured, it is effective to use a light source that emits light having a wavelength that strongly attracts the insects to be captured for the attracting light of the insect capturing apparatus. For this purpose, first, it is necessary to identify the worms that gather at each factory, etc., but since the identification work has been carried out exclusively manually, the work cost has been increased, and the reduction has been required. . In addition, since the work is a human work by visual observation, it is not only time consuming, but it is difficult to expect an objective and consistent identification result, and improvement has been demanded.

  In order to meet such demands, the applicant of the present application can first and foremost automatically and quickly identify insects that gather at each factory, etc., and at a relatively low cost. A method for identifying captured insects capable of obtaining a sexual identification result has been proposed (Patent Document 1: Japanese Patent No. 5690856).

Japanese Patent No. 5690856 Japanese Patent No. 2694726 Japanese Patent No. 2599101

  The identification method according to the above proposal includes an image reading step of reading an image of an adhesive sheet of an insect trapping device that has captured insects, an insect candidate region acquiring step of acquiring an insect candidate region by acquiring an edge from the image, A background region deleting step for deleting a background region other than the insect candidate region; an insect region extracting step for extracting an insect region from the insect candidate region; and extracting at least a body region, a leg region and a cocoon region from the extracted insect region And calculating the body length, body width and the ratio of body length to body width from the extracted body region, calculating the leg length and the ratio of leg length to body length from the leg region, and also calculating the heel area and heel area from the heel region The feature extraction step of calculating the ratio of the body area to obtain the Merckmar data of the morphological features, and the standard Mellmar data set and stored in advance for each insect It is intended to include the identification step of identifying the insects by collating the Mar data.

The identification method according to this proposal can satisfy the request to automatically and promptly identify insects that gather at each factory etc. at a relatively low cost. In this method, There are problems such as the processing of the white line area existing in the sheet image, which is an obstacle to the extraction of the object area, or the minute insect having an area area of 0.25 mm ² or less is judged as dust, There was a need for further improvements.

  The present invention has been made to meet such a demand. In the insect candidate region acquisition step, a white line region existing in a sheet image that becomes an obstacle to the extraction of an object region is processed and a process of extracting a low density region is performed. Including the opacity calculation and body color similarity in the morphological features of Merckmar data, in the identification step, the identification process using SVM is performed, and the dust determination process and the microworm determination process are performed. Thus, it is an object of the present invention to provide a method and an identification system for a captured reptile that can more accurately and quickly identify a worm, in particular, a minute worm at a relatively low cost.

  The invention according to claim 1 for solving the above-described problem includes an image reading step of an adhesive sheet of an insect trapping device that captures reptiles, an insect candidate region acquiring step of acquiring an insect candidate region from the read image, A background region deleting step for deleting a background region other than the acquired insect candidate region, an insect region extracting step for extracting an insect region from the insect candidate region, and Merckmar data of morphological features from the extracted insect region. A capture insect identification method including an acquired feature extraction step and an identification step of identifying the insect by comparing the Merckmar data with standard Merckal data.

  The insect candidate area acquisition step converts the RGB value into the HSV color system and uses the saturation and brightness to remove the white line area existing in the sheet image that becomes an obstacle to the extraction of the insect candidate area. Then, after extracting the white line area candidates, the process of extracting the white line area using the Hough transform, and normalizing so that the average of the pixels obtained with the focus on the target pixel in the insect candidate area becomes 100 is performed. And a process of extracting a low density region from a result calculated using a Gaussian operator.

The invention according to claim 2 for solving the above-described problem includes an image reading step of an adhesive sheet of an insect trapping device that captures reptiles, an insect candidate region acquiring step of acquiring an insect candidate region from the read image, A background region deleting step for deleting a background region other than the acquired insect candidate region, an insect region extracting step for extracting an insect region from the insect candidate region, and Merckmar data of morphological features from the extracted insect region. A method for identifying a captured reptile comprising: a feature extraction step to obtain; and an identification step for identifying the reptile by comparing the Merckmar data with standard Merckmar data,
The identification step is a method for identifying a capture insect, characterized by including an identification process performed using SVM.

  In one embodiment, the identification step further includes dust determination processing for acquiring dust candidates from the black ratio of the object region or the luminance value of the body region and determining dust from the roundness of the body and the circularity of the body region. .

  Further, in one embodiment, as the Merckmar data of the morphological feature, the opacity of the eyelid determined from the comparison between the hue value of the pixel in the coordinates of the eyelid region and the average hue value of the background region, and the coordinates of the body region When the difference between the hue value of the pixel and the average hue value of the heel region is small, the pixel of the coordinates of the body region is set as a similar color pixel, and extracted from the comparison between the similar color pixel and the total number of pixels of the body region Body color similarity.

Further, in one embodiment, the Merckmar data of the morphological feature includes a fine region length that is a feature amount indicating how many thin regions in the insect region exist, and the extraction of the fine region length includes: It goes through the following steps.
-Insect area enhancement process-Insect area extraction process-Fine area extraction process-Fine area length acquisition process

  The identification method and identification system for captured reptiles according to the present invention are as described above. The identification of captured reptiles is performed by acquiring Merckal data of morphological features from captured reptile images without human intervention. Because this can be done by image processing that compares with standard Merckmar data set for each insect in advance, identification of insects gathering at each factory etc. can be performed automatically and quickly at a relatively low cost. In addition, there is an effect that an objective and consistent identification result can be obtained.

  In particular, in the case of the method according to the present invention, in the image reading step, the process of removing the white line region existing in the sheet image that becomes an obstacle to the extraction of the object region is performed, and in the insect candidate region acquiring step, the low concentration region is extracted. In the identification step, identification processing is performed using SVM, dust candidates are obtained from the black ratio of the object region or the luminance value of the body region, and the dust is determined from the roundness of the body and the circularity of the body region. By performing this determination processing, there is an effect that enables more accurate and reliable identification of reptiles.

It is a simplified block diagram of the system for enforcing the identification method of the capture insects based on this invention. It is a perspective view which shows the structural example of the insect trapping apparatus used in the identification method of the trapping insects based on this invention. It is a flowchart which shows the rough flow of the identification method of the capture insects based on this invention. It is a figure which shows the insect candidate area | region acquisition method in the identification method of the capture insects based on this invention. It is a figure which shows the extraction method of the white line area | region in the identification method of the capture insects based on this invention. It is a figure which shows the extraction method of the low concentration area | region in the identification method of the capture insects based on this invention. It is a figure which shows the deletion method of the background area | region in the identification method of the capture insects based on this invention. It is a figure which shows the extraction method of the insect area | region in the identification method of the capture insects based on this invention. It is a figure which shows the calculation method of the cocoon opacity in the identification method of the capture insects based on this invention. It is a figure which shows the calculation method of the body color similarity in the identification method of the capture insects based on this invention. It is a flowchart which shows the flow of the identification step in the identification method of the capture insects which concerns on this invention. It is a figure for demonstrating the tertiary identification process of the identification step in the identification method of the capture insects based on this invention. It is a figure which shows the flow of the identification process using SVM in the identification method of the capture insects based on this invention. It is a figure for demonstrating the parameter adjustment method of the model of the identification process using SVM in the identification method of the capture insects which concerns on this invention. It is a figure for demonstrating the evaluation experiment method of the identification process using SVM in the identification method of the capture insects based on this invention. It is a figure which shows the body region extraction method of the yellow thrips in the identification method of the capture insects based on this invention. It is a graph which shows the hue example of a thrips and garbage in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of the black thrips in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of the black thrips in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of the black thrips in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of the thrips in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of chironomid in the identification method of the capture insects based on this invention. It is a figure which shows the determination method of the dust with the uniform color change in the identification method of the capture insects based on this invention. It is a figure which shows the extraction method of the unevenness | corrugation degree of a body region in the identification method of the capture insects based on this invention. It is a figure which shows the extraction method (extract method of an insect area | region) of the fine area length in the identification method of the capture insects which concerns on this invention. It is a figure which shows the extraction method (extraction method of a thin area | region) of the thin area length in the identification method of the capture insects which concerns on this invention. It is a figure which shows the extraction method (acquisition method of thin area length) of the thin area length in the identification method of the capture insects based on this invention.

Embodiments for carrying out the present invention will be described in detail with reference to the drawings. The method for identifying a capture insect according to the present invention includes:
-An image reading step of the adhesive sheet of the insect trapping device that captured the insects;
An insect candidate region acquisition step of acquiring an insect candidate region from the read image;
A background region deletion step of deleting a background region other than the acquired insect candidate region;
An insect region extraction step for extracting an insect region from the insect candidate region;
A feature extraction step of obtaining morphological feature Merckmar data from the extracted insect region;
-An identification step for identifying said reptiles by comparing said Merckmar data with standard Merckmar data.

And the capture insect identification method according to the present invention further includes the following processes.
-In order to remove the white line area existing in the sheet image that becomes an obstacle to the extraction of the object area in the image reading step, the RGB value is converted into the HSV color system, and the white line area is used by using the saturation and the lightness. White line region extraction processing for extracting a white line region using Hough transform after extracting candidates-In the insect candidate region acquisition step, the average of pixels obtained centering on the target pixel in the object region is 100 After performing normalization, low density area extraction processing is performed to extract low density areas from the results calculated using the Gaussian operator

  Furthermore, in the identification step, an identification process using SVM is performed, and dust candidates are acquired from the black ratio of the object region or the luminance value of the body region, and the dust is determined from the roundness of the body and the circularity of the body region. A dust determination process for performing the determination process is performed.

  Further, the identification method of the trapping insects according to the present invention is, as Merckmar data of morphological features, opacity of cocoons determined from a comparison between the hue value of the pixel of the coordinates of the cocoon region and the average hue value of the background region, and When the difference between the hue value of the pixel of the body region coordinate and the average hue value of the heel region is small, the pixel of the body region coordinate is a similar color pixel, and the total number of pixels of the similar color pixel and the body region It includes the body color similarity extracted from the comparison.

  FIG. 1 is a simplified block diagram of a system for carrying out a method for identifying a captured insect according to the present invention. The system captures an insect by an adhesive sheet 1 and adheres to insects recovered from the insect trap 1. The image reading means 2 for reading the image of the adhesive sheet, the analysis device 3 for analyzing the image read by the image reading means 2 and identifying the captured insects, and the result identified by the analysis device 3 are output. And output means 8.

  The analysis device 3 includes a preprocessing unit 4 that performs image processing for distinguishing the insect region and the background region on the image read by the image reading unit 2, and the morphological features of the insects from the preprocessed image. Feature extraction means 5 for extracting the Merckmar data, data storage means 6 for storing the standard Merckmar data obtained by standardizing the Merckmar data of the morphological features of the reptiles, the extracted Merckmar data and the data storage extracted by the feature extraction means 5 It is configured to include identification means 7 for identifying the worm by collating with the standard Merckmar data stored in the means 6 (see FIG. 1).

  Further, as the insect trapping device 1 for catching insects, for example, a frame-shaped main body case 11 having a front opening and a front opening formed therein, and a light guide plate insertion opening 16 on the side surface of the main body case 11 are installed. A light guide plate provided with an LED unit 13 and a light-transmitting pressure-sensitive adhesive sheet 15 that is superposed on the light guide plate and loaded into the main body case 11 is used (see FIG. 2). In that case, the adhesive sheet 15 is arranged so that the adhesive surface is directly exposed in the front opening 12 of the main body case 11 when the main body case 11 is loaded.

  In the present invention, the insects to be identified include chironomid mosquitoes, ganganbo, butterflies, flea flies, drosophila, black fly mushrooms, mosquito flies, wing flies, house flies, fly flies, leafhoppers, hornets, ants, moths, moths, cutlet worms , Kokunu Tomodoki, Garbushi, Hanakakushi, Aphid, Aokamushi, Marukamushi, Leafhopper, Leafhopper, Chatterbug, Thrips, and Mayfly.

Each of these reptiles has morphological features, but the morphological features can be classified to some extent. For example, the morphological characteristics of the house fly, gangabo, squid, chironomid, butterfly, flea fly, scallop fly, black-winged fly, and fake fly are as follows.
-Ganbo, squid, chironomid, black spring mushroom ... large body (10 mm or more).
-Giganbo, Musca, Chironomid, Flea, Tamabae ... Long legs and developed.
-Mushroom flies, fake flies ... The tactile sensation is thick and short and has a bumpy shape.
-Butterfly fly, Flyfly ... The wings are wide and symmetrical.
-Hyme flies, gangambo, fleas, black-winged flies, fake flies ... has characteristic veins.

  In view of these morphological characteristics, as the morphological characteristics of Merckmar, insect body length, body width, body length / body width ratio, body color, leg length, leg length / body length ratio, tail length, tail length / body length ratio, heel area, The area of the heel / body area, the body part (which part of the body), the angle of the heel, the direction of the head, the state of the back leg, the degree of roundness of the body, the state of the eye (position, aspect ratio), etc. Conceivable. Usually, in the present invention, at least the body length, body width, leg length, leg length / body length ratio, heel area, heel area / body area ratio are employed. In addition, the identification of the reptiles in the present invention is performed by observing the captured reptiles in a state of being adhered to the adhesive sheet, not in a normal state. It is also considered that there are cases where sticking is captured.

  Furthermore, in the present invention, as the Merckmar data of the morphological feature, the opacity of the eyelid determined from the comparison between the hue value of the pixel of the eyelid region coordinate and the average hue value of the background region, and the pixel of the body region coordinate When the difference between the hue value and the average hue value of the eyelid region is small, the pixel of the coordinates of the body region is set as a similar color pixel, and the body color extracted from the comparison between the similar color pixel and the total number of pixels of the body region Including color similarity. Hail opacity is data used to identify similar forms of insects based on their transparency, and body color similarity is used to identify similar forms of insects based on differences in body color This data is used for

  Hereinafter, the method according to the present invention will be described in detail for each processing step with reference to the flowchart shown in FIG. Note that the flowchart of FIG. 3 shows steps from the processing by the image reading means 2, which is an insect identification operation, from analyzing the read image to identifying it and outputting the identification result.

Standard Merckmar data creation and storage step This step creates a standard Merckmar data (acceptable range) by setting the morphological features of the worms for each insect as a template prior to the identification of the reptiles. It is a step of storing in the means 6.

Image reading step (S11)
In this step, the image is read from the insect collecting device 1 installed in a factory or the like, and the adhesive sheet 15 to which the insects adhered to the film are wrapped as necessary is applied to the image reading device (image input device) 2 to read the image. It is.

Insect candidate region acquisition step (S12)
In the insect candidate area acquisition step, the edge 31 is acquired from the read image (hereinafter referred to as insect image) (FIG. 4A) (FIG. 4B), and the insect image is divided into blocks (for example, 1 The range of the block is 200 × 200 pixels), and a block including the edge 31 is acquired and labeling is performed (FIG. 4C). The acquisition of the edge 31 can be performed by a general edge detection method.

White Line Region Extraction Processing The sheet image of the pressure-sensitive adhesive sheet 15 has a white line region (FIG. 5A) that is used when the worker visually counts. Since the white line region is an obstacle to extracting the insect candidate region by image processing, it is necessary to extract the white line region when extracting the insect candidate region. The extraction of the white line region can be performed using the Hough transform. For this purpose, first, white line candidate regions are extracted. Since the white line has a feature that the lightness is high, the RGB value can be converted into the HSV color system by the following formula 1 and extracted using the saturation (S) and the lightness (V) (FIG. 5B). )).

Next, a white line region is extracted from the white line candidate region. For this purpose, first, the extracted white line candidate region is subjected to a Hough transform in the vertical direction to obtain the difference between the x-coordinates of the upper and lower two points. Hereinafter, the difference between the x coordinates is set to D (FIG. 5C). Then, the upper left corner of the image is set to coordinates (0, 0), the lower left corner is set to coordinates (0, Bottom), the upper right corner is set to coordinates (Right, 0), the lower right corner is set to coordinates (Right, Bottom), and the coordinates of point A are set to ( X _A , 0) is moved pixel by pixel in the x direction from (0,0) to (Right, 0). At the same time when the point A is moved, the point B of the coordinates (X _A ± D, Bottom) is set (FIG. 5D). Next, a straight line connecting points A and B is acquired, and when the straight line passes a certain area or more on the white line area, a white line candidate area in a 10 × 10 range on the straight line is extracted as a white line area (FIG. 5). (F)). After the point A reaches (Right, 0), the point A is returned to (0, 0).

Subsequently, the coordinate of the point A is set to (0, Y _A ), and is moved pixel by pixel in the y direction toward (0, Bottom). At the same time when the point A is moved, the point B of coordinates (Right, Y _A ± D) is set. Next, a straight line connecting points A and B is acquired, and when the straight line passes a certain area or more on the white line area, a white line candidate area in a 10 × 10 range on the straight line is extracted as a white line area (FIG. 5). (H)). When the point A reaches (0, Bottom), the extraction of the white line region is completed (FIG. 5 (i)).

Low concentration region extraction processing Also, in order to facilitate the extraction of insect candidate regions, processing for extracting low concentration regions from the insect candidate regions is performed. For this purpose, first, a density in a range of 5 × 5 is extracted centering on the target pixel. FIG. 6A shows the insect candidate area and the target pixel, and FIG. 6F shows the extracted 5 × 5 area. Next, normalization is performed so that the average of the extracted 25 (5 × 5) pixels is 100. FIG. 6G shows the normalized values. For the normalized pixel, a value is calculated using a Gaussian operator weighted at the center. FIG. 6 (i) shows the operator, and FIG. 6 (b) shows the result of calculating the value. As a result of performing the same calculation for all the pixels, a pixel having a low output value is calculated as a low density region. FIG. 6E shows a low concentration region.

Since the extraction of the low density region is for determining whether or not the low density region is in the 5 × 5 range in this way, even if the pixel in which the noise is generated becomes low density, Is not determined, and is determined to be a low density region only when the low density exists over a wide range.
The low density region ratio can be used for determination of insects having a black pattern or insects having a characteristic that a part of the body is black.

Background region deletion step (S13)
In order to extract the characteristics of the captured insects, it is necessary to extract the insect area from the insect candidate area 32 (see FIG. 7A), and in order to extract the insect area, the background area is extracted from the insect candidate area 32. Need to be deleted. For this purpose, first, the HSV color information of the non-edge block area 33 not including the edge 31 is acquired, and then the edge block area including the edge 31, that is, from the bug candidate area 32 and the non-edge block area 33, the non-edge block area The same color area as that of the 33 HSV color information is considered as a background and deleted (FIG. 7B).

  However, the insect area obtained by simply removing the background in this way includes the shadow of the insect in addition to the original insect area. Can't hope. Therefore, in the present invention, the insect region extracted as described above is not definitive, and is assumed to be a worm region 34 that is presumed to be a worm region. Image processing for confirmation is performed.

Insect region extraction step (S14)
In this step, first, edge detection is performed on the pseudoworm area 34 extracted using the color information by a general edge detection method (FIG. 8A). In FIG. 8, the dark portion indicated by reference numeral 35 is an edge. Next, the pixel of interest 36 in the gradient direction is extracted (positioned inside the edge 35 as shown in FIG. 8B), and for each pixel adjacent to each pixel of interest 36, four neighbors adjacent to it. If the difference is within a certain value (for example, abs (integrated pixel luminance value−target pixel luminance value) <15 (luminance value: 0 to 255)), the pixel is focused on. The luminance of the pixel 36 is integrated.

Morphological feature extraction step (S15)
After the insect region 38 is extracted as described above, the body region, the leg region, the heel region, and the tail region are first extracted based on the extracted insect region 38. This morphological feature extraction step includes a body region, a leg region, a heel region, and a tail region extraction step, and a morphological feature Merckmar data acquisition step based thereon.

  The body region is a region other than the worm's legs, moths, and tails. To extract the body region, first, HSV color information of the insect region 38 and the background region 39 is acquired. Then, an area having a color different from that of the background area 39 is extracted as a body area.

  Needless to say, the leg region is a much narrower and narrower region than the body region. Therefore, a process of fitting a 4 × 4 square block to the insect area is performed, and a narrow part sandwiched between the edges is extracted as a narrow area, and the 4 × 4 square block is not fitted, A portion extracted as a narrow region is set as a leg candidate region. The leg region is extracted by tracing the thin line from the end point of the thin line in the leg candidate region to the body region. That is, the thinning process is performed on the leg candidate region, and the leg region is defined from the end point of the thin line in the leg candidate region to the end point of the thin line in the body region.

  The cocoon region is extracted by labeling a region excluding the leg candidate region from the insect image, and setting the maximum region among them as a cocoon region. In addition, the tail region is extracted by performing a process of fitting a 4 × 4 square block to the body region as in the case of extracting the leg candidate region and extracting a narrow portion sandwiched between the edges as a narrow region. A portion where the 4 × 4 square block does not fit and a portion where the portion extracted as the narrow region overlaps are defined as the tail region.

  Subsequently, Merckmar data of morphological features is acquired from the body region, leg region, heel region, and tail region extracted as described above. That is, the body length and body width are calculated from the body region, the leg length and the leg length to body length ratio are calculated from the leg region, and the heel area and the heel area to body area ratio are calculated from the heel region. Other morphological features of Merckmar include body color, tail length, tail length / length ratio, slaughtered part, heel angle, head direction, back leg condition, body roundness, eye position and aspect ratio, etc. In addition, particularly in the present invention, it includes wrinkle opacity and body color similarity. Any one of these can be used in combination. Of course, the morphological feature of Merckmar is not limited to these. Further, the feature extraction method is not limited to the method described below.

<Extraction of body length, body width and body length / body width ratio>
The body length is extracted by detecting the skeleton in the body region and measuring the length. The detection of the skeleton and the measurement of the length are performed by the following steps.
-Find the center of gravity of the body.
-Extract one straight line from the center of gravity of the body part to the most distal point.
-Extract one second straight line up to the most distal point at a different angle from the first straight line.
-Consider the first straight line and the second straight line as a skeleton, and measure their lengths.

  As for the body width, a plurality of straight lines orthogonal to the first straight line and the second straight line are extracted, and the length of the longest straight line among the plurality of orthogonal straight lines is defined as the body width. Then, the body length / body width ratio (body length / body width ratio) is extracted as one of the features from the body width value thus obtained and the body length value obtained as described above.

<Extraction of body color>
In extracting the characteristics of the body color, first, four color regions of black, red, yellow, and green are extracted from the body region using the HSV color system. In the extraction of the four-color region, first, for black, for example, when 0 ≦ V value <80 is considered as a threshold value, for a portion not corresponding to this, for example, red: 0 <H value <40, yellow: 40 <H Value <70, green: 70 <H value <100 is used as a threshold value, and the ratio, saturation average, and brightness average of each color are acquired and extracted as body color features.

<Extraction of leg length and body length / leg length ratio>
The leg length can be obtained from the number of pixels of the thin line traced from the end point in the leg region to the end point in the body region. The length ratio between the body length and the leg length can be obtained from the leg length / body length.

<Extraction of tail length and body length / tail length ratio>
The tail length can be obtained from the length of the longest straight line that can be created in the tail region. The length ratio between the body length and the tail length can be obtained from the tail length / body length.

<Extraction of heel area and body area / heel area ratio>
The heel area can be obtained from the area of the heel region, and the ratio of the body area to the heel area can be obtained from the heel area / body area.

<Extraction of cocoon>
The extraction of the cocoon part is to detect from which part of the body part the cocoon comes out. For the extraction, the body part is divided into a front part, an intermediate part, and a rear part, and any part that is in contact with the heel is defined as a slaughtered part.

<Calculation of heel angle>
The heel angle is calculated as follows. -Obtain a straight line from the center of the skeleton to the tip of the heel. -Get two corners between a straight line and a skeleton. -The smaller one of the two obtained angles is calculated as the angle of the eyelid.

<Calculation of opacity>
In the morphological feature extraction step, the opacity of the cocoon can be mentioned as the Merckmar data of the morphological feature.抽出 Opacity extraction is performed when a pixel at coordinates (i, j) is opaque when the hue value H _{i, j} of the pixel at coordinates (i, j) in the 翅 region is significantly different from the average hue value H _back of the background region. This is done by extracting as pixels. When the opaque pixels occupy more than half of the total number of pixels in the cocoon region, it is determined that there is opacity, and the ratio of opaque pixels is calculated. Equation 2 shows the conditions for opaque pixels, FIG. 9A shows a worm image, FIG. 9B shows a cocoon region (portion surrounded by a thick line), and FIG. 9C shows an opaque pixel.

<Calculation of body color similarity>
Furthermore, the similarity of the color of the body and the eyelid can be mentioned as the Merckmar data of the morphological features. Extraction of the body and wing color similarity coordinate the body region (i, j) the hue value H _i of _pixels, when the difference between the average hue value H _WinG of _j and翅領region is small, the coordinates (i , J) are extracted as similar color pixels. When the similar color pixels occupy more than half of the total number of pixels in the body region, it is determined that there is a body similar color, and the ratio of similar color pixels is calculated. Equation 3 shows the conditions for opaque pixels. FIG. 10A shows an insect image, FIG. 10B shows a cocoon region, a body region (a portion surrounded by a thick line), and FIG. 10C shows similar color pixels.

<Estimation of insect head direction>
The head direction of the insect can be estimated based on the angle of the moth determined as described above. That is, when the angle of the heel is small, the direction opposite to the heel is the head direction, and when the angle of the heel is large, the thicker one (the body width is large) of the front part and the rear part of the body part Is the head direction.

<Extraction of back legs>
The back leg is the back leg with the leg extending from the direction opposite to the head direction.

<Feature extraction of body roundness>
The feature of the roundness of the body part can be extracted from an average value of slopes at both ends of the histogram by obtaining an approximate quadratic function from the histogram of the body width.

<Eye feature extraction>
In order to extract eye features, first, a black part that is a candidate for an eye is extracted by binarizing the body part. Then, the black portion is identified as an eye by a portion having both the requirement to be located at the distal end portion of the skeleton and the portion having a shape satisfying, for example, the aspect ratio <2.0.

<Extraction of narrow area length>
The fine region length is a feature amount that represents how many thin regions in the insect region exist. The extraction of the fine region length is performed through the following steps.
-Insect area enhancement process-Insect area extraction process-Fine area extraction process-Fine area length acquisition process

Insect region enhancement step This step is a step of enhancing the insect region in order to facilitate extraction of only insects from the image. The image is composed of RGB channels. Here, attention is paid only to the G channel. This is because the G channel has a large color difference between the sheet and the insect.

Insect region extraction step This step is a step of extracting the insect region from the image emphasized in the enhancement step, and extracts only the insect region based on the difference from the background color. For this purpose, first, the mode of luminance is obtained from the image, and then a raster operation is performed to extract a region where the pixel of interest satisfies Equation 4 as a bug candidate region. The threshold value of Equation 4 is obtained from Equation 5.
Attention pixel value <mode of the background region−threshold (Formula 4)
Threshold = −7 / 15 × (average value of object area) +78 (Formula 5)
Threshold = 8 when threshold <8
When 64 <threshold, threshold = 64
After that, labeling is performed, and the largest area is extracted as an insect area (see FIG. 25).

Extraction step This step fine region is a step of extracting a narrow area from insects region. In this step, first, the contraction process is performed twice and the expansion process is performed twice on the insect region image (see FIG. 26A) to obtain an opening image (see FIG. 26B). Then, after performing a top hat process to extract a thin region, only a region having an area of 2.5% or more of the insect region is extracted, and the extracted region is set as a thin region (see FIG. 26C).

Step of obtaining the narrow region This step is a step for obtaining the length of the narrow region. For this purpose, first, a closing image is obtained by performing expansion processing twice and contraction processing twice on the fine region (FIG. 27A) (see FIG. 27B). Next, a thinning process is performed on the closing image, and the number of pixels remaining after the thinning process is set as a thin area length (see FIG. 27C).

Identification step (S16)
Next, an identification method using the feature data extracted by the above method will be described with reference to the flowchart shown in FIG. The identification step includes a classification process for classifying the identification target insects into three types, harsh, sterile, and giant (S21), an identification process ranging from primary to quaternary (S22, S24, S26, S28), and identification results Result output step (S23, S25, S27, S29).

The classification into the three types of 翅, 翅, and giant is done to improve the efficiency of processing. First, it is determined whether or not it is a giant reptile, and the insect identified as not a giant reptile Identify the presence or absence of a class. Whether or not it is huge is identified by, for example, whether or not the condition of the entire area> 10 mm ² is satisfied. Usually, gangambo and hiratabu are classified as giant reptiles. Next, for the reptiles that are not identified as giant reptiles, for example, it is identified whether or not the condition of the body / moth area ratio> 0.1 is satisfied. If it does not meet the criteria, it is classified into an insect-free species (for example, Knotnutomodoki, Katsuobushimushi, Hanekakushi, Shibamushi).

In the primary identification step (S22), for each reptile, the numerical range (threshold value) of the Merckmar data is set, and it is identified whether the feature of the identification target insect corresponds to the threshold value of any reptile, When it corresponds, the process which identifies the identification object insect with the said insect is made | formed. As an example of the threshold value, in the case of a trapped chironomid (small black) with high capture probability, body length (mm) 0.50-2.00, body width (mm) 0.20-0.85), body length / The body width ratio is 0.90 to 1.60, the heel area (mm ² ) is 0.10 to 1.00, and the body / heel area ratio is 0.10 to 1.30. When identification is possible at the stage of the primary identification step (S22), the identification result (insect name) is output (S23), and when the identification is not possible, the process proceeds to the secondary identification step (S24).

  The secondary identification step (S24) is a step of expanding the threshold value and identifying whether or not an identification target insect that does not correspond to the primary identification step (S22) corresponds to the expanded threshold value. In this case, it is preferable to extend the threshold from an insect having a high capture probability. If there is a corresponding item in the secondary identification step (S24), the identification target insect is identified as the insect, and the result is output (S25). Moreover, when there is no applicable thing, it transfers to a tertiary identification process (S26).

  The tertiary identification step (S26) is a final determination step of the insect name, and as sub-steps there are multiple types (for example, body length, body width, body length / body width ratio, leg length, leg length / body length ratio, heel area, heel area) -9 types of body area ratio, area, and body area), a step of setting a representative value, a step of calculating a Euclidean distance between the representative value and the insect to be identified, and three types in the order of the shortest distance Determining identification candidates. As the representative value, an average value of a plurality of samples is adopted.

  In the example shown in FIG. 12, the representative values are 1.5 mm body length and 0.5 mm leg length in the case of Kurobane mushroom fly, 3.0 mm body length and 0.0 mm leg length in the case of Drosophila, and the body length in the case of flies. It is 1.0 mm, leg length is 1.5 mm, the chironomid case is set as 0.5 mm body length and leg length 0.5 mm, and the same is true when the body length of the insect to be identified is 1.5 mm and the leg length is 0.5 mm. This is a regular example. In this case, when the Euclidean distance between these representative values and the reptile to be identified is 0.70 for D. melanogaster, 1.58 for Drosophila, 1.11 for Drosophila, 1.00 for chironomid, In order from the first place, the order is Kurobane mushroom fly, chironomid, gall fly, and Drosophila.

The name of the insect is finally identified from the three types of identification candidates determined. Usually, the first candidate is used as the identification result (in the above example, the insect to be identified is identified as the black fly mushroom fly). In this case, however, it is possible to move up the candidate ranking for a reptile having a unique characteristic on the condition of the characteristic. For example, fleas can be raised to the first candidate when three or more of the following four characteristics are satisfied.
-The percentage of yellow for each body color is 40% or more-The rear legs are extracted-The cocoon is in the middle-Black eyes are extracted

  As described above, in the method according to the present invention, instead of identifying the reptile from the silhouette of the whole reptile, first, each part of the body region, the leg region and the cocoon region is extracted, and the insect is determined from the characteristics of each part. Is used. Thereby, it becomes possible to reduce the influence by the difference in the direction of the insect trapped by sticking as much as possible. In addition, the problem of the presence of similar types of insects with similar colors, shapes, and sizes can be solved by increasing the number of features in each part.

In the tertiary identification step (S26), the determination is performed mainly using the similarity of the color feature amount, but the colors are similar to each other, for example, chironomid and flea, Drosophila and leafhopper unka. Cannot distinguish insects. Therefore, in order to further improve the determination accuracy when it is determined as a specific insect in the tertiary identification step, it is preferable to perform identification using SVM (Support Vector Machine) (fourth identification step (S28)). The insects that are identified in the quaternary identification step (S28) are, for example, as follows.
・ Identify flies or chironomies in the third identification step (S26) when it is determined to be fleas ・ Identify Drosophila or leafhoppers in the third identification step (S26) if it is determined to be Drosophila or leafhopper

  The identification using the SVM includes a learning data preparation step (S31), a learning data normalization step (S32), a model selection step (S33), a model parameter adjustment step (S34), and an evaluation experiment step ( S35) (see FIG. 13).

Learning data preparation step (S31)
The learning data preparation step (S31) is a step of preparing learning data as a sample of classification. It is preferable to collect as much learning data as possible. In the present invention, SVM classification is performed in 47 dimensions using the following 47 feature values.

The following features may include features that are not effective for identification. However, as the characteristics of SVM, data that was not selected is not incorporated into the discrimination function, so many features are used. There is no problem.
・ Body color (15)
Red ratio, red saturation, red brightness, yellow ratio, yellow saturation, yellow brightness, green ratio, green saturation, green brightness, black ratio, blackness, own ratio, whiteness, black ratio (RGB), low density ratio Body area (11)
Area, body length, body width, body width dispersion, body length body width ratio, circularity, axis angle, roundness, luminance average, luminance dispersion, thin region length, object region (1)
Black ratio ・ Tail area (2)
Length / length ratio-Wrinkle area (9)
Total area / Total area ratio / Maximum area / Maximum area ratio / Maximum length / Body length ratio / Opacity ratio / Color similarity ratio-3 areas (4)
Area, luminance average, luminance dispersion, circularity, leg area (2)
Length / body length ratio • 2nd body region (2)
Area / contrast ratio

Normalization process of learning data (S32)
The learning data normalization step (S32) is a step of performing a process of previously adjusting a possible range of the feature amount in order to prevent the weight for each feature amount from being changed. For example, in the case of the body region area (1.24 to 3.65) and the body region luminance variance (260 to 990), the value variation of the body region luminance variance is large, and if normalization is not performed, the result is greatly affected. I will do it. Also, normalization of learning data is necessary to prevent information loss from occurring in the kernel function. Normalization of the learning data is performed using the following formula 6.

Model selection process (S33)
In the model selection step (S33), a kernel function and model parameters are selected. Kernel functions include RBF kernel, linear kernel, polynomial kernel, sigmoid kernel, etc., but it is recommended to use the most basic RBF kernel. The parameters for the RBF kernel are the cost parameter C and the parameter γ of the RBF kernel.

Model parameter adjustment step (S34)
The model parameter adjustment step (S34) is a step of tuning to increase the accuracy of the parameters. The cost parameter C is a parameter that determines how much misclassification is allowed, and the parameter γ of the RBF kernel is a parameter that determines whether the decision boundary is simple or complex. These parameters are adjusted by performing a grid search to obtain an optimal combination of the cost parameter C and the RBF kernel parameter γ. That is, as shown in FIG. 14, the two-dimensional space of C and γ is divided into grids, and what seems to be optimal at the intersection is obtained. For this purpose, first, a method is used in which a search is performed with a wider width to estimate the optimum location, and then a search is performed in detail around the value that becomes highly accurate.

Evaluation experiment process (S35)
In the evaluation experiment step (S35), the cross-validation method (k-fold) is used to rearrange the data and perform the accuracy evaluation a plurality of times. This method has the advantage that the accuracy of the estimation accuracy is small and the reliability of accuracy can be maintained even when the amount of data is small. FIG. 15 shows an example in which 3fold cross-validation is performed with 6 original data by the cross-validation method.

An experimental example through the quaternary identification step (S28) using the SVM is shown below.
・ Original data: A total of 144 Drosophila and 72 leafhoppers ・ Model + parameter RBF function (C: 512, γ: 0.00048828125)
・ Accuracy evaluation 36fold cross validation ・ Experimental result Estimated accuracy: 68.06% (98/144)

  Preferably, in the identification step (S16), a dust determination process and a microworm determination process are further performed.

Very small insects, such as the determination process of garbage "thrips" and "midges" is erroneously determined garbage, there is a problem that can not be extracted. In addition, a piece of an insect leg may be erroneously determined as an insect. Therefore, in the present invention, these problems are solved by adding dust determination processing to the identification step (S16). A major characteristic of the dust attached to the sheet is that it has a distorted shape. Further, since the body is not round as compared with insects and the object color is a single color of black or white, dust is determined using these characteristics.

  First, the ratio of black in the object region is obtained, and the candidate is determined to be a dust candidate when the ratio is 95% or more or the luminance value of the body region is 180 or more. Next, the roundness of the body and the circularity of the body region are obtained with respect to the dust candidates, and when the values are very small, it is determined that the shape is distorted and determined as dust.

Until now, if the area of the area is 0.25 mm ² or less, it has been determined as dust. Therefore, there is a problem that very small insects such as thrips are also determined as dust. Therefore, it was decided to solve these problems by adding judgment processing for very small insects (Thrips aurium) and insects with a small body area (Chironomidae). First, a method for determining thrips will be described. There are two types of thrips, yellow thrips and black thrips, and it is necessary to determine these two types of thrips.

Determination of yellow thrips A yellow thrips object region is labeled to extract a region of 50 pixels or less, and a body candidate region is extracted from that region. The candidate area is obtained by applying blocks of 2 × 2 pixels to the object area in order from the upper left, and letting all applicable areas be body candidate areas. A region satisfying (Equation 7.1) in this body candidate region is extracted as a body region. At this time, when the extracted body region occupies 30% or more of the object region, the hue variation of the object region is obtained, and when the variation is large, it is determined as thrips, and when the variation is small, it is determined as dust. 16 (a) shows a thrips, FIG. 16 (b) shows a block area, FIG. 16 (c) shows a body candidate area, FIG. 16 (d) shows a body area, and FIG. 17 (a) shows a thrips hue. For example, FIG. 17B shows an example of the hue of dust.

Judgment of black thrips There are three types of black thrips, each with different characteristics. The first type is thin at the center and black around it. The second type has an overall black character. The third type is mixed with red and has a black area. Therefore, the determination is performed by performing separate processing for each thrips. The first type of thrips is characterized by a thin center and black surroundings (FIG. 18 (a)). To this end, first, V-value color information is acquired from the HSV color system for the background area, and an average value V _ave1 is obtained. Next, the V value of the acquired object region is compared with V _ave1 , the pixel of the object region that satisfies the condition (formula 7.2.1) is extracted as a dark region, and labeling is performed (FIG. 18B). . At this time, a thin area is extracted from (Equation 7.2.2) using the RGB values of the object area (FIG. 18C).

Determination of yellow thrips Next, a block of 5 × 5 pixels is applied to a thin area, and when there are two or more dark areas in the block (FIG. 19), determination is made as thrips. The second type of thrips has the characteristic of being black overall (FIG. 20 (a)). In the same manner as the extraction of the first type of thrips, a dark region is extracted from Equation 6.2.1 and labeled (FIG. 20B). At this time, if the extracted dark area occupies 40% or more of the object area, it is determined as thrips. The third type of thrips is mixed with red and has a black area (FIG. 21A). Therefore, first, color information of H and V values is acquired from the HSV color system for the background region, and average values H _ave1 and V _ave1 are obtained. Next, the H and V values of the acquired object region are compared with H _ave1 and V _ave1 , and pixels in the object region that satisfy the condition (Equation 7.2.3) are extracted as regions having different hues. At this time, the pixels of the object region that satisfy the condition (Formula 7.2.4) are extracted as dark regions. Finally, if there is a dark area above a certain level and an area with a different hue occupies 40% or more of the object area, it is determined as thrips (FIG. 21B).

The chironomid judgment method Next, the chironomid judgment method is explained. In the previous research, there was a problem that the body area of small chironomid was not extracted correctly and was judged as garbage. Therefore, in the present invention, the determination process is performed using the fact that the body area, which is a characteristic of the small chironomid, is small, and the tail is green, thereby solving the problem of erroneous determination of dust. For this purpose, first, a green area is obtained from (Equation 7.2.5) from the RGB values of the object region (FIG. 22B). Next, when the obtained green area is 25% or more of the body area and the body region pixels are 80 pixels or less, it is considered that there is a possibility of chironomid. Finally, the chironomi feature width is set, and when the feature quantity to be identified matches the set condition, it is determined to be chironomid.

Determination of dust with uniform color change As shown in FIG. 23B, dust is determined with little color change. The size of the target candidate area is set to 0.4 mm ² ≤ bug candidate area ≤ 1.0 mm ² . First, Log-Polar conversion is performed with a size of 64 × 64 with the insect candidate region as the center of gravity. Then, the standard deviation σ of the luminance value for each row is calculated. Thus, the magnitude of the color change from the center toward the outside is compared. If the standard deviation of all the rows is 15 or less, it is determined as dust. The reason for developing from the center is to reduce the influence of the direction of the region.

The extraction area of the unevenness is thinned, and a line segment connecting the start point and the end point of the thinning result is drawn. And it can obtain | require by length other than the body area | region on a line segment / length of a line segment (FIG. 24).

  As described above, in the method for identifying a captured insect according to the present invention, in the insect candidate region acquisition step, the white line region present in the sheet image that is an obstacle to the extraction of the object region is processed and the low concentration region is extracted. In the identification step, the identification process performed using SVM is performed in the identification step, and the dust determination process and the microworm determination process are performed. By doing so, the identification of reptiles, especially the identification of micro-insects, can be carried out more accurately and quickly, and at a relatively low cost, so that the industrial applicability is great.

DESCRIPTION OF SYMBOLS 1 Insect capture device 2 Image reading means 3 Analysis apparatus 4 Pre-processing means 5 Feature extraction means 6 Data storage means 7 Identification means 8 Output means 11 Main body case 12 Front opening part 13 LED unit 15 Translucent adhesive sheet 16 Light guide plate insertion port 31 Edge 32 Insect candidate area
33 non-edge block area 34 pseudoworm area 35 edge 36 pixel of interest 37 adjacent pixel 38 insect area 39 background area

Location This invention relates to the identification how capture insects, and more particularly, for example, food factories, as such chemicals plant, you need to minimize blocking the entry of insects and other small insects in, identified the captured insects, it relates to the identification how to capture insects in order to obtain the materials to make the subsequent insect measures.

The present invention has been made to meet such a demand. In the insect candidate region acquisition step, a white line region existing in a sheet image that becomes an obstacle to the extraction of an object region is processed and a process of extracting a low density region is performed. Including the opacity calculation and body color similarity in the morphological features of Merckmar data, in the identification step, the identification process using SVM is performed, and the dust determination process and the microworm determination process are performed. in the identification of insects, particularly the identification of small insects more accurately and quickly, yet it is an object to provide a relatively identification how capture insects that can be performed at a low cost.

In one embodiment, the identification step includes an identification process performed using SVM.

Identification how capture insects according to the present invention has been as described above, the identification of trapping insects, without human intervention, to get the Merkmal data morphological features from the image of the captured insects, which Can be performed by image processing to compare with standard Merckmar data set for each insect in advance, so that identification work of insects gathering at each factory etc. can be performed automatically and quickly at a relatively low cost. In addition, there is an effect that an objective and consistent identification result can be obtained.

Claims (5)

An image reading step of the adhesive sheet of the insect trapping device that captured the insects, an insect candidate region acquiring step for acquiring an insect candidate region from the read image, and a background for deleting a background region other than the acquired insect candidate region A region deletion step, a bug region extraction step for extracting a bug region from the bug candidate region, a feature extraction step for obtaining morphological feature Merckmar data from the extracted bug region, and the Merckmar data as standard Merckmar data A method for identifying a captured reptile comprising an identification step of identifying the reptile by matching,
The insect candidate area acquisition step converts the RGB value into the HSV color system and removes the white line area existing in the sheet image that becomes an obstacle to the extraction of the insect candidate area, and uses the saturation and the brightness. After extracting a white line area candidate, a process of extracting a white line area using Hough transform,
Processing for extracting a low density region from a result calculated using a Gaussian operator after normalizing so that an average of pixels obtained with the pixel of interest in the insect candidate region as a center is 100. A method for identifying captured insects characterized by An image reading step of the adhesive sheet of the insect trapping device that captured the insects, an insect candidate region acquiring step for acquiring an insect candidate region from the read image, and a background for deleting a background region other than the acquired insect candidate region A region deletion step, a bug region extraction step for extracting a bug region from the bug candidate region, and a feature extraction step for obtaining Merckmar data of morphological features from the extracted bug region;
An identification step of identifying a captured reptile, including an identification step of identifying the reptile by comparing the Merckmar data with standard Merckmar data,
The method for identifying a captured insect, wherein the identification step includes an identification process performed using SVM.   The identification step further includes dust determination processing for acquiring dust candidates from the black ratio of the object region or the luminance value of the body region, and determining dust from the roundness of the body and the circularity of the body region. To identify captive insects.   As the Merckmar data of the morphological features, the opacity of the eyelid determined from the comparison between the hue value of the pixel at the coordinates of the eyelid area and the average hue value of the background area, the hue value of the pixel at the coordinates of the body area, and the eyelid area When the difference from the average hue value of the body region is small, the pixel of the coordinates of the body region is a similar color pixel, and the body color similarity extracted from the comparison between the similar color pixel and the total number of pixels of the body region The identification method of the capture insects in any one of Claims 1 thru | or 3 containing. As the Merckmar data of the morphological features, including a fine region length which is a feature amount indicating how many thin regions in the insect region exist, the fine region length is extracted through the following steps, The method for identifying a capture insect according to any one of claims 1 to 4.
-Insect area enhancement process-Insect area extraction process-Fine area extraction process-Fine area length acquisition process