CN111627059A - Method for positioning center point position of cotton blade - Google Patents

Abstract

The invention discloses a method for locating the center point of a cotton leaf, comprising the following steps: S1: collecting a cotton image; S2: extracting the leaf contour of each cotton leaf in the image; S3: extracting four coordinates of the leaf contour of a certain cotton leaf Points x1, x2, x3 and x4; S4: Calculate the coordinates of the initial center point of the leaf profile of the cotton blade; S5: Obtain the abscissa x _c of the center point; S6: Obtain the ordinate y _c of the center point; S7: The coordinates (x _c, y _c ) are the position of the center point of the cotton blade. The method of the invention has a small amount of calculation, so that it has good real-time performance, and the center accuracy meets the requirements of practical applications. In the real-time test, the real-time processing speed index FPS value is 20.107; in terms of positioning accuracy, the test finally obtains the average error The MAE is 26.861, which meets the requirements for positioning accuracy in practical application scenarios.

Description

A method for locating the center point of cotton leaves

technical field

The invention relates to the technical field of automatic identification and positioning of crops, in particular to a method for positioning the center point of cotton leaves.

Background technique

The center point positioning method of cotton leaves based on image processing can quickly obtain the coordinate position of the center point of the target crop, and provide a reference for the precise spraying and automatic management of the target crop in the field. This method can improve the efficiency of refined crop management and reduce the dependence on additional labor. and optimal control of agricultural costs. Obtaining the image of the target crop through a specific camera has the advantages of low cost, high flexibility and high efficiency. At present, there are also methods such as laser sensor, infrared sensor, depth camera, etc. to locate the coordinates of the center point of the image (irregular geometric image), but these methods are generally There are shortcomings such as large amount of calculation, high production cost, and poor real-time performance. Therefore, the method of positioning the center point of cotton leaves based on image processing is more practical in actual production and life.

The cotton leaf center point location method based on image processing is to use the camera to obtain the real image information of the target crop, and through a series of image analysis and calculation processing, the location of the center point of the target crop leaf center is finally realized. At present, the mainstream leaf center positioning algorithm mainly obtains the coordinates of the crop center point by extracting the contour of the crop and the connected area, and calculating the centroid of the connected area. However, this method has high requirements on shape rules, and has poor real-time performance and adaptability to images such as irregular geometric shapes, complex backgrounds, and overlapping target objects.

In order to make up for the influence of positioning errors caused by the irregular shape of crops and the complex background, the existing technology uses the image histogram target crop statistics to determine the crops to be tested, and then performs histogram statistics on the target crop rows to determine the target crops. (Wang Yongkang et al. An indoor positioning method based on image grayscale histogram similarity calculation [J]. Bulletin of Surveying and Mapping, 2018(4):63-67.). However, due to the common overlapping phenomenon of actual target crop leaves, this method has a large center point positioning error in specific application scenarios.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a method for locating the position of the cotton leaf center point, which has better real-time performance and whose center point positioning accuracy meets the application requirements of the actual scene.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions:

A method for locating the center point of a cotton leaf, comprising the following steps:

S1: Collect cotton images;

S2: extract the leaf contour of each cotton leaf in the image;

S3: Extract four coordinate points of the leaf profile of a cotton leaf, which are the upper left coordinate point x1 (x _ltop , y _ltop ), the upper right coordinate point x2 (x _rtop , y _rtop ), and the lower left coordinate point x3 (x _llower , y _llower ) and the lower right coordinate point x4(x _rlower ,y _rlower );

S4: Calculate the coordinates (x ₀ , y ₀ ) of the initial center point of the leaf profile of the cotton leaf according to the following formula:

S5: Take the initial center point as the starting point, set the horizontal step size △x, x _llower ≤△x≤x _rlower , take the vertical axis passing through the initial center point as the main axis, and perform horizontal scanning to the left and right to obtain the horizontal direction of the center point. coordinate x _c ;

S6: Take the initial center point as the starting point, set the vertical step length △y, y _llower ≤△y≤y _ltop , and take the horizontal axis passing through the initial center point as the main axis to perform vertical scanning up and down, and obtain the ordinate of the center point y _c ;

S7: The coordinates (x _c , y _c ) are the position of the center point of the cotton blade.

In step S1 of the above method, the cotton image is collected and uploaded using the prior art. In step S2, the existing technology is used to extract the leaf contour of each cotton leaf in the image (that is, the leaf area of the cotton leaf is segmented), such as using the MaskR-CNN target detection model to perform cotton leaf segmentation on the cotton image, based on the convolutional neural network. The target leaf segmentation and transfer learning-based image segmentation, etc.; the leaf contours of the extracted cotton leaves are usually rectangular or square.

In step S5 of the above method, the lateral step size Δx can be set as required, and in general, the smaller the step size Δx, the more accurate the positioning of the center point. Since it needs to take the longitudinal axis passing through the initial center point as the main axis to perform lateral scanning to the left (that is, to the negative direction of the x-axis with the initial center point as the origin) and to the right (that is, to the positive direction of the x-axis with the initial center point as the origin), so There are positive and negative values for the horizontal step size △x; when performing horizontal scanning to the left, the horizontal step size △x is a negative value, and when performing horizontal scanning to the right, the horizontal step size △x is a positive value. In a preferred embodiment of the present application, the lateral step size Δx is set to ±5px. The step S5 further includes:

S501, initialize the number of initial scans i=1, and the vertical axis passes through the coordinates (x ₀ ^* =x ₀ +△x,y ₀ =0);

S502 , during scanning, the blade contour is divided into two image areas through the vertical axis of the coordinates (x ₀ ^* , y ₀ ), which are the image area ABCD and the image area A ^* B ^* C ^* D ^* respectively, where:

The four vertex positions of the image area ABCD are: A(x _ltop ,y _ltop ), B(x ₀ +△x,y _ltop ), C(x ₀ +△x,y _llower ), D(x _llower ,y _llower );

The four vertex positions of the image area A ^* B ^* C ^* D ^* are: A ^* (x ₀ +△x,y _rtop ), B ^* (x _rtop ,y _rtop ), C ^* (x _rlower ,y _rlower ) , D ^* (x ₀ +△x,y _llower );

S503, the image areas ABCD and A ^* B ^* C ^* D ^* are respectively scaled to images of the same specification, and further converted into grayscale images;

In step S503, the image areas ABCD and A ^* B ^* C ^* D ^* are scaled and scaled to the same specification, so as to retain basic information such as structure, light and shade, and discard the comparison differences caused by images of different sizes/ratios. The scaling requirements of the image areas ABCD and A ^* B ^* C ^* D ^* are set as required, usually scaling to (20～30)×(20～30) size, more preferably scaling to 20px×20px , i.e. each image is 400 pixels. Converting the scaled image into a grayscale image can further reduce the amount of redundant information in the image and improve the real-time processing efficiency of the algorithm.

S504, calculate the mean and median of the pixels in each row of the two grayscale images respectively, and record the mean and median of the pixels in each row;

S505, calculate the standard deviation of all means and medians of the two graphs obtained in step S504, respectively, to obtain the mean standard deviation and median standard deviation (the mean value standard after the image area ABCD is scaled and further converted into a grayscale image) The difference and median standard deviation are denoted as σ _avg and σ _median , respectively, and the mean standard deviation and median standard deviation after the image area A ^* B ^* C ^* D ^* is scaled and further converted into a grayscale image are denoted as σ , respectively _avg ^* and _σmedian ^* ) as their respective numerical features, namely:

Numerical feature of image area ABCD: ABCD=(σ _avg ,σ _median );

Numerical features of image area A ^* B ^* C ^* D ^* : A ^* B ^* C ^* D ^* = (σ _avg ^* ,σ _median ^* );

S506, using the numerical features of the image areas ABCD, A ^* B ^* C ^* D ^* , and calculating the similarity imageSimilarity of the two through the cosine function; that is,

S507, save the calculated similarity and the abscissa x ₀ ^* of the vertical axis in the current number of scans into the result list imageSimilaritys in the form of a tuple, that is, imageSimilaritys.append((x ₀ ^* ,imageSimilarity)), end the current scan scanning;

S508, update the x ₀ ^* coordinates, that is, redefine the sum of x ₀ ^* in the i-th scan and the lateral step size △x as x ₀ ^* in the i+1-th scan, and judge that the redefined x ₀ ^* and The relationship between x _rlower and x _llower , when x ₀ ^* >x _rlower or x ₀ ^* <x _llower , end the right or left lateral scan with the vertical axis as the main axis, and execute step S509; otherwise, execute step S502 to carry out the i-th +1 scan until the updated x ₀ ^* satisfies the condition for ending the right or left lateral scan with the vertical axis as the main axis;

S509, after finishing the rightward or leftward horizontal scanning with the vertical axis as the main axis, start the leftward or rightward horizontal scanning with the vertical axis as the main axis. The vertical axis is the main axis for the left and right horizontal scan results list imageSimilaritys;

S510, traverse the result list imageSimilaritys, find out the tuple data with the smallest similarity value, and use x ₀ ^* in the tuple data as the abscissa x _c of the cotton leaf center point.

In step S6 of the above method, the vertical step Δy can be set as required, and the value of the vertical step Δy and the horizontal step Δx can be the same or different, and preferably the same. Since it is necessary to take the horizontal axis passing through the initial center point as the main axis to perform vertical scanning upward (that is, taking the initial center point as the origin to the positive direction of the y-axis) and downward (that is, taking the initial center point as the origin to the y-axis direction), the longitudinal scan is The step size △y also has positive and negative values; when the vertical scanning is performed upward, the vertical step size △y is a negative value, and when the vertical scanning is performed downward, the vertical step size △y is a positive value. The step S6 further includes:

S601, initialize the number of initial scans i'=1, and the horizontal axis passes through the coordinates (x ₀ =0, y ₀ ^* =y ₀ +△y);

S602 , during scanning, the blade contour is divided into two image areas through the horizontal axis of the coordinates (x ₀ , y ₀ ^* ), which are image area A'B'C'D' and image area A' ^* B' ^* C respectively ' ^* D' ^* , where:

The four vertex positions of the image area A'B'C'D' are: A'((x _ltop , y _ltop )), B'(x _rtop ,y _rtop ), C'(x _rtop ,y ₀ +△ y), D'(x _ltop ,y ₀ +△y);

The four vertex positions of the image area A' ^* B' ^* C' ^* D' ^* are: A' ^* (x _llower ,y ₀ +△y), B' ^* (x _llower ,y ₀ +△y), C' ^* (x _rlower , y _rlower ), D' ^* (x _llower , y _llower );

S603: Scale the image areas A'B'C'D' and A' ^* B' ^* C' ^* D' ^* respectively, and zoom to the vertical axis passing through the initial center point as the main axis to perform horizontal scanning to the left and right In the process, the image areas ABCD and A ^* B ^* C ^* D ^* are scaled to the same specifications, and are further converted into grayscale images;

In step S603, the purpose of scaling the image areas A'B'C'D' and A' ^* B' ^* C' ^* D' ^* is the same as the aforementioned step S503.

S604, calculate the mean and median of the pixels in each row of the two grayscale images respectively, and record the mean and median of the pixels in each row;

S605, calculate the standard deviation of all the mean and median of the two images obtained in step S604 respectively (the standard deviation of the mean and the median after the image area A'B'C'D' is scaled and further converted into a grayscale image The differences are expressed as σ _avg ' and σ _median ' respectively, and the mean standard deviation and median standard deviation of the image area A' ^* B' ^* C' ^* D' ^* after scaling and further conversion into grayscale images are respectively expressed as σ _avg ' ^* ,σ _median ' ^* ), take the obtained mean standard deviation and median standard deviation as their respective numerical features, namely:

Numerical features of image area A'B'C'D': A'B'C'D'=(σ _avg ',σ _median ');

Numerical features of image area A' ^* B' ^* C' ^* D' ^* : A' ^* B' ^* C' ^* D' ^* = (σ _avg ' ^* ,σ _median ' ^* );

S606, using the numerical features of the image areas A'B'C'D', A' ^* B' ^* C' ^* D' ^* , calculate the similarity imageSimilarity' of the two through the cosine function; that is,

S607, save the calculated similarity and the ordinate y ₀ ^* that the horizontal axis passes through in the current number of scans to the result list imageSimilaritys' in the form of a tuple, that is, imageSimilaritys'.append((y ₀ ^* , imageSimilarity')), end the current scan;

S608, update y ₀ ^* coordinates, that is, redefine the sum of y ₀ ^* in the i'th scan and the vertical step Δy as y ₀ ^* in the i'+1th scan, and determine the redefined y ₀ ^* Relationship with y _ltop or y _llower , when y ₀ ^* >y _ltop or y ₀ ^* <y _llower , end the upward or downward vertical scanning with the horizontal axis as the main axis, and execute step S609; otherwise, execute step S602 for the first i'+1 scans until the updated y ₀ ^* satisfies the condition for ending the upward or downward vertical scan with the horizontal axis as the main axis;

S609, after finishing the upward or downward vertical scanning with the horizontal axis as the main axis, start the downward or upward vertical scanning with the horizontal axis as the main axis, the implementation steps are the same as S601-S608, and the horizontal axis including the initial center point at the same time is obtained. It is the result list imageSimilaritys' of the upward and downward vertical scanning of the main axis;

S610, traverse the result list imageSimilaritys', find out the tuple data with the smallest similarity value, and use y ₀ ^* in the tuple data as the ordinate y _c of the center point of the cotton leaf.

In this application, the transverse direction is the horizontal direction, the longitudinal direction is the longitudinal direction perpendicular to the horizontal direction, the transverse axis is the horizontal axis parallel to the horizontal line, and the longitudinal axis is the longitudinal axis perpendicular to the horizontal line .

Compared with the prior art, the method of the present invention has a small amount of calculation, which makes the real-time performance better, and the accuracy of the center meets the practical application requirements. The performance of the real-time processing speed index FPS value is 20.107; in terms of positioning accuracy, the artificial standard picture and the positioning error marked by the algorithm are used for testing (considering the interference of the actual environment, the allowable positioning coordinate error is 0 ~ 5px, then The acceptable average error MAE in the actual scene is in the range of 0 to 25). The final average error MAE obtained in the test is 26.861, which is close to the upper error limit of 25, which meets the requirements of the actual application scene for positioning accuracy.

Description of drawings

Fig. 1 is the flow chart of the cotton blade center point position location method of the present invention;

Fig. 2 is the cotton image obtained (that is, the cotton image to be uploaded);

Fig. 3 is the result graph after the cotton leaf area is divided on the image shown in Fig. 2 (that is, the result graph after extracting the leaf contour of each cotton leaf in the image shown in Fig. 2);

Fig. 4 is from Fig. 3 and intercepts the leaf profile picture of a certain cotton leaf independently;

FIG. 5 is a schematic diagram of the initial center point position (x ₀ , y ₀ ) of the cotton leaf (hereinafter also referred to as the target crop) in the picture shown in FIG. 4 ;

FIG. 6 is a schematic diagram of performing horizontal scanning to the left and right with the vertical axis as the main axis taking the initial center point position (x ₀ , y ₀ ) of the target crop as the starting point;

Fig. 7 is the area segmentation effect diagram after moving +5px to the right (that is, in the positive direction of the x-axis) with the vertical axis as the main axis taking the initial center point position (x ₀ , y ₀ ) of the target crop as the starting point;

8 is a schematic diagram of vertical scanning upward and downward with the initial center point position (x ₀ , y ₀ ) of the target crop as the starting point and the horizontal axis as the main axis;

Fig. 9 is a region segmentation effect diagram after moving +5px to the y-axis upward (that is, to the positive direction of the y-axis) with the horizontal axis as the main axis, taking the initial center point position (x ₀ , y ₀ ) of the target crop as the starting point;

Figure 10 is a schematic diagram of the location of the center point of the target crop in a theoretical environment;

Figure 11 is an overview of the positioning effect of the center point position of target crops of different sizes in the actual environment.

Detailed ways

The present invention will be further described in detail below in conjunction with the specific drawings to better understand the content of the present invention, but the present invention is not limited to the following examples.

Referring to Fig. 1, the method for locating the position of the cotton blade center point of the present invention comprises the following steps:

S1: Collect cotton images;

S2: extract the leaf contour of each cotton leaf in the image;

S3: Extract four coordinate points of the leaf profile of a cotton leaf, which are the upper left coordinate point x1 (x _ltop , y _ltop ), the upper right coordinate point x2 (x _rtop , y _rtop ), and the lower left coordinate point x3 (x _llower , y _llower ) and the lower right coordinate point x4(x _rlower ,y _rlower );

S4: Calculate the coordinates (x ₀ , y ₀ ) of the initial center point of the leaf profile of the cotton leaf according to the following formula:

S5: Take the initial center point as the starting point, set the horizontal step size △x, x _llower ≤△x≤x _rlower , take the vertical axis passing through the initial center point as the main axis, and perform horizontal scanning to the left and right to obtain the horizontal direction of the center point. coordinate x _c ;

S6: Take the initial center point as the starting point, set the vertical step length △y, y _llower ≤△y≤y _ltop , and take the horizontal axis passing through the initial center point as the main axis to perform vertical scanning up and down, and obtain the ordinate of the center point y _c ;

S7: The coordinates (x _c , y _c ) are the position of the center point of the cotton blade.

In a specific embodiment, the method of the present invention is described in detail by taking the image shown in FIG. 2 as the acquired cotton image.

S1: Take the image shown in Figure 2 as the acquired cotton image and upload it.

S2: Use the existing Mask R-CNN target detection model to perform cotton leaf segmentation on the cotton image shown in Figure 2, and the result is shown in Figure 3. The leaf outline of the extracted cotton leaves is rectangular or square.

S3: Take the leaf profile picture of a certain cotton leaf in FIG. 3 separately (as shown in FIG. 4 ), and extract four coordinate points of the leaf profile of the cotton leaf at the same time, which are respectively recorded as the upper left coordinate point x1 (x _ltop , y _{ltop )} ), the upper right coordinate point x2(x _rtop ,y _rtop ), the lower left coordinate point x3(x _llower ,y _llower ), and the lower right coordinate point x4(x _rlower ,y _rlower ).

S4: Calculate the coordinates (x ₀ y ₀ ) of the initial center point of the leaf profile of the cotton leaf according to the following formula (as shown in FIG. 5 ):

S5: Take the initial center point as the starting point, set the horizontal step size △x=±5px, and use the vertical axis passing through the initial center point as the main axis to scan left and right, and obtain the abscissa x _c of the center point (as shown in the figure) 6), specifically, the vertical axis is used as the main axis to scan to the right for display, including:

S501, initialize the number of initial scans i=1, and the vertical axis passes through the coordinates (x ₀ ^* =x ₀ +5px, y ₀ =0);

S502 , during scanning, the blade contour is divided into two image areas by the vertical axis of the coordinates (x ₀ ^* , y ₀ ), namely the image area ABCD and the image area A ^* B ^* C ^* D ^* (as shown in FIG. 7 ). ),in:

The four vertex positions of the image area ABCD are: A(x _ltop ,y _ltop ), B(x ₀ +5px,y _ltop ), C(x ₀ +5px,y _llower ), D(x _llower ,y _llower ) ;

The four vertex positions of the image area A ^* B ^* C ^* D ^* are: A ^* (x ₀ +5px, y _rtop ), B ^* (x _rtop , y _rtop ), C ^* (x _rlower , y _rlower ), D ^* (x ₀ +5px,y _llower );

S503, scale the image areas ABCD and A ^* B ^* C ^* D ^* to 20px×20px specifications respectively (that is, each image is 400 pixels), and further convert them into grayscale images by using the existing technology;

S504, calculate the mean and median of the pixels in each row of the two grayscale images respectively, and record the mean and median of the pixels in each row;

S505, calculate the standard deviation of all means and medians of the two graphs obtained in step S504, respectively, to obtain the mean standard deviation and median standard deviation (the mean value standard after the image area ABCD is scaled and further converted into a grayscale image) The difference and median standard deviation are denoted as σ _avg and σ _median , respectively, and the mean standard deviation and median standard deviation after the image area A ^* B ^* C ^* D ^* is scaled and further converted into a grayscale image are denoted as σ , respectively _avg ^* and _σmedian ^* ) as their respective numerical features, namely:

Numerical feature of image area ABCD: ABCD=(σ _avg ,σ _median );

Numerical features of image area A ^* B ^* C ^* D ^* : A ^* B ^* C ^* D ^* = (σ _avg ^* ,σ _median ^* );

S506, using the numerical features of the image areas ABCD, A ^* B ^* C ^* D ^* , and calculating the similarity imageSimilarity of the two through the cosine function; that is,

以元组形式保存至结果列表imageSimilaritys中，即imageSimilaritys.append((

imageSimilarity))，结束当前次的扫描；S507, the calculated similarity and the abscissa passed by the vertical axis in the current number of scans

Save it to the result list imageSimilaritys in the form of a tuple, that is, imageSimilaritys.append((

imageSimilarity)), end the current scan;

S508, update the x ₀ ^* coordinates, that is, redefine the sum of x ₀ ^* (x ₀ ^* = x ₀ +5px) and +5px in the first scan as x ₀ ^* in the second scan, and determine that the redefinition is required The relationship between x ₀ ^* and x _rlower and x _llower , when x ₀ ^* >x _rlower or x ₀ ^* <x _llower , end the rightward horizontal scan with the vertical axis as the main axis, and execute step S509; otherwise, execute step S502 to carry out The second scan, until the updated x ₀ ^* meets the conditions for ending the rightward horizontal scan with the vertical axis as the main axis;

S509, since it is not only necessary to perform leftward horizontal scanning with the longitudinal axis passing through the initial center point as the main axis, but also rightward transverse scanning with the vertical axis passing through the initial center point as the main axis, therefore, after the end of the vertical axis as the main axis After the right horizontal scan, it is necessary to start the left horizontal scan with the vertical axis as the main axis. The implementation steps of the left horizontal scan are the same as S501~S508, except that the value of the horizontal step is negative (-5px), until the vertical axis is completed. It is the leftward horizontal scan of the main axis, and the result list imageSimilaritys that contains both the left and right horizontal scans with the vertical axis as the main axis is obtained.

S510, traverse the result list imageSimilaritys, find out the tuple data with the smallest similarity value, and use x ₀ ^* in the tuple data as the abscissa x _c of the cotton leaf center point.

S6: Take the initial center point as the starting point, set the vertical step size △y=±5px, and use the horizontal axis passing through the initial center point as the main axis to perform vertical scanning up and down, and obtain the vertical coordinate y _c of the center point (as shown in Figure 8). shown); specifically, the horizontal axis is used as the main axis to perform an upward vertical scan for display, including:

S601, initialize the number of initial scans i'=1, and the horizontal axis passes through the coordinates (x ₀ =0, y ₀ ^* =y ₀ +5px);

S602 , during scanning, the blade contour is divided into two image areas through the horizontal axis of the coordinates (x ₀ , y ₀ ^* ), which are image area A'B'C'D' and image area A' ^* B' ^* C respectively ' ^* D' ^* (as shown in Figure 9), where:

The positions of the four vertices of the image area A'B'C'D' are: A'((x _ltop , y _ltop )), B'(x _rtop , y _rtop ), C'(x _rtop , y ₀ +5px ), D'(x _ltop ,y ₀ +5px);

The four vertex positions of the image area A' ^* B' ^* C' ^* D' ^* are: A' ^* (x _llower , y ₀ +5px), B' ^* (x _llower , y ₀ +5px), C' ^* (x _rlower , y _rlower ), D' ^* (x _llower , y _llower );

S603, scale the image areas A'B'C'D' and A' ^* B' ^* C' ^* D' ^* to 20px×20px images respectively, and further convert them into grayscale images by using the existing technology;

S604, calculate the mean and median of the pixels in each row of the two grayscale images respectively, and record the mean and median of the pixels in each row;

S605, calculate the standard deviation of all the mean and median of the two images obtained in step S604 respectively (the standard deviation of the mean and the median after the image area A'B'C'D' is scaled and further converted into a grayscale image The differences are expressed as σ _avg ' and σ _median ' respectively, and the mean standard deviation and median standard deviation of the image area A' ^* B' ^* C' ^* D' ^* after scaling and further conversion into grayscale images are respectively expressed as σ _avg ' ^* ,σ _median ' ^* ), take the obtained mean standard deviation and median standard deviation as their respective numerical features, namely:

Numerical features of image area A'B'C'D': A'B'C'D'=(σ _avg ',σ _median ');

Numerical features of image area A' ^* B' ^* C' ^* D' ^* : A' ^* B' ^* C' ^* D' ^* = (σ _avg ' ^* ,σ _median ' ^* );

S606, using the numerical features of the image areas A'B'C'D', A' ^* B' ^* C' ^* D' ^* , calculate the similarity imageSimilarity' of the two through the cosine function; that is,

S607, save the calculated similarity and the ordinate y ₀ ^* that the horizontal axis passes through in the current number of scans to the result list imageSimilaritys' in the form of a tuple, that is, imageSimilaritys'.append((y ₀ ^* , imageSimilarity')), end the current scan;

S608, update y ₀ ^* coordinates, that is, redefine the sum of y ₀ ^* (y ₀ ^* =y ₀ +5px) in the first scan and the vertical step size △y as y ₀ ^* in the second scan, and judge The redefined relationship between y ₀ ^* and y _ltop or y _llower , when y ₀ ^* >y _ltop , ends the upward vertical scan with the horizontal axis as the main axis, and executes step S609; otherwise, executes step S602 for the second scan, Until the updated y ₀ ^* satisfies the condition for ending the upward vertical scan with the horizontal axis as the main axis;

S609, since not only the upward vertical scanning with the horizontal axis passing through the initial center point as the main axis is required, but also the downward vertical scanning with the horizontal axis passing through the initial center point as the main axis, after the end of the upward vertical scanning with the horizontal axis as the main axis After scanning, it is necessary to start the downward vertical scanning with the horizontal axis as the main axis. The implementation steps of the downward vertical scanning are the same as S601 to S608, except that the value of the vertical step is a negative value (-5px) until the horizontal axis is completed. It is the downward vertical scan of the main axis, and the result list imageSimilaritys' that includes both the upward and downward vertical scans of the horizontal axis as the main axis is obtained.

S610, traverse the result list imageSimilaritys', find out the tuple data with the smallest similarity value, and use y ₀ ^* in the tuple data as the ordinate y _c of the center point of the cotton leaf.

S7: The coordinates (x _c , y _c ) are the position of the center point of the cotton blade (as shown in FIG. 10 ).

Figure 11 shows a list of effects of using the method of the present invention to locate the center points of target crops of different sizes in an actual environment.

The number of frames per second (PFS) of the picture in the above-mentioned specific embodiment is tested, and the test method and the result are as follows:

The FPS test method is as follows:

Step 1: Use the file_read function to read all the target crop access addresses in the system directory, and save them in the seeds in the form of a list;

Step 2: Complete the extraction of seeds 100, 500, 900, 1300, 1800 scales by slicing according to the seeds in Step 1;

Step 3: By looping through the seeds100, seeds500, seeds900, seeds1300, seeds1800 seed files, call the algorithm one by one to process each image and record the time used by the algorithm to process a single image each time. The total time at the seed scale, and the average time per image is obtained by dividing the total time by the total number of pictures. Since the processing time of all single images is recorded, the record of the minimum time and the highest time of a single image can be finally obtained;

Step 4: The final "aggregate/average single time", or FPS, is obtained by accumulating the average single time at different scales and dividing by the total number of scales.

In order to verify the real-time processing performance of the algorithm, a test simulation is carried out on a self-built image database (a total of 1820 target crops), and Table 1 lists the test results of the real-time processing speed of the algorithm of the present invention on different scale data volumes, The time unit is milliseconds (ms).

Table 1:

Test the positioning accuracy of the center point in the above-mentioned specific embodiment:

The center point positioning accuracy verification mainly uses 100 images marked by manual positioning in advance as the benchmark. At the same time, the above images are marked with secondary positioning through the algorithm, and the coordinate points dot _origin (x _origin , y _origin ) and The Euclidean distance between the later coordinate points dot _predict (x _predict , y _predict ) represents the error between the two, and finally the average positioning error is obtained by dividing the accumulated error value by 100.

Positioning accuracy test method:

Step 1: Obtain 100 target crop pictures, and manually mark the coordinates of the center point for each picture;

Step 2: The algorithm performs secondary positioning and labeling on the 100 manually labeled target crop images;

Step 3: Calculate the Euclidean distance between the coordinate point dot _origin (x _origin , y _origin ) of the manual positioning annotation and the identified coordinate point dot _predict (x _predict , y _predict );

Step 4: Accumulate the errors in step 3 to obtain the cumulative sum of errors of all pictures, and divide the cumulative sum of errors by 100 to obtain the average positioning error of a single picture, ie MAE. The results are shown in Table 2 below.

Table 2:

Claims (4)

1. a method for positioning the center point position of a cotton blade is characterized by comprising the following steps:

s1: collecting cotton images;

s2: extracting the leaf outline of each cotton leaf in the image;

s3: extracting four coordinate points of the blade outline of a certain cotton blade, namely an upper left coordinate point x1 (x)_ltop,y_ltop) Upper right coordinate point x2 (x)_rtop,y_rtop) Lower left coordinate point x3 (x)_llower,y_llower) And a lower right coordinate point x4 (x)_rlower,y_rlower)；

S4: calculating the coordinate (x) of the initial center point of the blade profile of the cotton blade according to the following formula₀,y₀)：

S5, setting lateral step length △ x, x with the initial center point as a starting point_llower≤△x≤x_rlowerPerforming left and right transverse scanning by using a longitudinal axis passing through the initial central point as a main axis to obtain an abscissa x of the central point_c；

S6, setting longitudinal step length △ y, y with the initial center point as a starting point_llower≤△y≤y_ltopTaking a transverse axis passing through the initial central point as a main axis to carry out upward and downward longitudinal scanning to obtain a longitudinal coordinate y of the central point_c；

S7: coordinate (x)_c,y_c) Namely the position of the central point of the cotton blade.

2. The method as set forth in claim 1, wherein the step S5 further comprises:

s501, initializing the number of first scans i to 1, and passing the coordinate (x) on the ordinate₀ ^*＝x₀+△x,y₀＝0)；

S502, in scanning, the blade profile is passed through the coordinate (x)₀ ^*,y₀) Is divided into two image areas, namely an image area ABCD and an image areaDomain A^*B^*C^*D^*Wherein:

the four vertex positions of the image area ABCD are: a (x)_ltop,y_ltop)、B(x₀+△x,y_ltop)、C(x₀+△x,y_llower)、D(x_llower,y_llower)；

Image area A^*B^*C^*D^*The four vertex positions of (a) are respectively: a. the^*(x₀+△x,y_rtop)、B^*(x_rtop,y_rtop)、C^*(x_rlower,y_rlower)、D^*(x₀+△x,y_llower)；

S503, image areas ABCD and A^*B^*C^*D^*Respectively zooming to pictures with the same specification, and further converting the pictures into gray level images;

s504, respectively calculating the mean value and the median of the pixel points of each line of the two gray-scale images, and recording the mean value and the median of the pixel points of each line;

s505, calculating the standard deviations of all the mean values and the median values of the two graphs obtained in step S504, and using the obtained mean value standard deviations and the obtained median value standard deviations as the respective numerical characteristics, that is:

numerical characteristics of the image area ABCD: ABCD ═ σ (σ)_avg,σ_median)；

Image area A^*B^*C^*D^*The numerical characteristics of (A): a. the^*B^*C^*D^*＝(σ_avg ^*,σ_median ^*)；

S506, using the image areas ABCD, A^*B^*C^*D^*Calculating the similarity imageSimilarity of the two values by a cosine function;

s507, calculating the obtained similarity and an abscissa x passing through by a longitudinal axis in the current scanning times₀ ^*Saving in tuple form into the result list imageSimilaritys, i.e. imageSimilaritys₀ ^*imageSimilarity)), end of the current timeScanning;

s508, updating x₀ ^*Coordinates, i.e. x in the i-th scan₀ ^*Redefined as x in the i +1 th scan, as the sum of the transverse steps △ x₀ ^*Determining the redefined x₀ ^*And x_rlowerAnd x_llowerWhen x is₀ ^*>x_rlowerOr x₀ ^*<x_llowerIf so, the rightward or leftward horizontal scanning with the vertical axis as the main axis is finished, and step S509 is executed; otherwise, step S502 is executed to perform the (i + 1) th scanning until the updated x₀ ^*Satisfying the condition of ending the right or left transverse scanning taking the longitudinal axis as the main axis;

s509, after the right or left transverse scanning with the longitudinal axis as the main axis is finished, the left or right transverse scanning with the longitudinal axis as the main axis is started, the steps are the same as S501 to S508, and a result list imageSimiaritys which simultaneously contains the left or right transverse scanning with the longitudinal axis passing through the initial central point as the main axis is obtained;

s510, traversing the result list imageSimiaritys, finding out the tuple data with the minimum similarity value, and using x in the tuple data₀ ^*Abscissa x as center point of the cotton leaf_c。

3. The method as set forth in claim 1, wherein the step S6 further comprises:

s601, initializing the number of first scans i' to 1, and the horizontal axis passes through the coordinate (x)₀＝0,y₀ ^*＝y₀+△y)；

S602, scanning, the blade profile is passed through the coordinate (x)₀，y₀ ^*) The horizontal axis of (A) is divided into two image regions, image region A ' B ' C ' D ' and image region A '^*B’^*C’^*D’^*Wherein:

the four vertex positions of the image area a 'B' C 'D' are: a' ((x)_ltop，y_ltop))、B’(x_rtop,y_rtop)、C’(x_rtop,y₀+△y)、D’(x_ltop,y₀+△y)；

Image region A'^*B’^*C’^*D’^*The four vertex positions of (a) are respectively: a'^*(x_llower,y₀+△y)、B’^*(x_llower,y₀+△y)、C’^*(x_rlower,y_rlower)、D’^*(x_llower,y_llower)；

S603, image areas A ' B ' C ' D ' and A '^*B’^*C’^*D’^*Respectively zooming to the image areas ABCD and A in the process of performing horizontal scanning to the left and the right by taking a longitudinal axis passing through the initial central point as a main axis^*B^*C^*D^*The scaled specifications are the same and are further converted into a gray scale image;

s604, respectively calculating the mean value and the median of the pixel points of each line of the two gray-scale images, and recording the mean value and the median of the pixel points of each line;

s605, calculating the standard deviations of all the mean values and the median values of the two graphs obtained in step S604, and using the obtained mean standard deviations and median standard deviations as the respective numerical characteristics, that is:

numerical characteristics of image area a 'B' C 'D': a ' B ' C ' D ═ (σ)_avg’,σ_median’)；

Image region A'^*B’^*C’^*D’^*The numerical characteristics of (A): a'^*B’^*C’^*D’^*＝(σ_avg’^*,σ_median’^*)；

S606, image areas A ' B ' C ' D ' and A '^*B’^*C’^*D’^*Calculating the similarity imageSimilarity' of the two through a cosine function;

s607, the calculated similarity and the ordinate y through which the horizontal axis passes in the current scanning times₀ ^*Saving to result list in tuple formin imageSimilaritys ', i.e.' appind ((y)₀ ^*imageSimilarity')), ending the current scan;

s608, updating y₀ ^*Coordinates, i.e. y in the i' th scan₀ ^*The sum with the longitudinal step △ y is redefined as y in the i' +1 th scan₀ ^*Determining the redefined y₀ ^*And y_ltopOr y_llowerWhen y is₀ ^*>y_ltopOr y₀ ^*<y_llowerWhen the scanning is finished, the upward or downward longitudinal scanning with the horizontal axis as the main axis is finished, and step S609 is executed; otherwise, step S602 is executed to perform the i' +1 th scanning until the updated y₀ ^*The condition of ending the upward or downward longitudinal scanning by taking the transverse axis as the main axis is met;

s609, after finishing the upward or downward longitudinal scanning by taking the horizontal axis as the main axis, starting the downward or upward longitudinal scanning by taking the horizontal axis as the main axis to realize the steps from S601 to S608, and obtaining a result list imageSimiliarys' simultaneously containing the upward and downward longitudinal scanning by taking the horizontal axis passing through the initial central point as the main axis;

s610, traversing the result list imagesimilarity', finding out the tuple data with the minimum similarity value, and using y in the tuple data₀ ^*As the ordinate y of the center point of the cotton blade_c。

4. The method for locating the position of the center point of a cotton blade according to any one of claims 1 to 3, wherein the transverse step Δ x and the longitudinal step Δ y have the same or different values.

Images