CN103116747A - Method and system for automatically recognizing images of stalks and leaves of corns - Google Patents

Abstract

The invention discloses a method for recognizing stalks and leaves of corns. The method includes steps of (1), acquiring an RGB (red, green, blue) color image; (2), converting the RGB color image into a grayscale image; (3), preprocessing the grayscale image; (4), converting the preprocessed grayscale image into a binary image by means of binarization processing; (5), performing straight line detection for the binary image; and (6), computing shape features of different zones, and recognizing the acquired RGB color image according to the shape features of the different zones to finally acquire a stalk target. The different zones are acquired by means of detecting the binary image and are separated from one another by straight lines. The method has the advantages that stalks of the corns can be accurately recognized in real time, and problems of damaging equipment or hurting corn plants and the like due to blockage of stalks when an automatic fertilizer applicator is used for fertilization operation can be prevented.

Description

Method and system for automatically recognizing images of corn stems and leaves

technical field

The invention relates to the technical field of image processing, in particular to a method for identifying corn stalks.

Background technique

During the corn fertilization process, the fertilization device on the automatic fertilizer spreader will collide with "obstacles" such as corn stalks, resulting in equipment damage or injury to crops. Therefore, real-time and accurate identification and positioning of these "obstacles" is required , causing the automatic fertilizer applicator to change its movement route or stop before it may collide with the "obstacle".

This identification and positioning technology can be realized through machine vision, using cameras to collect images, and using computers, DSP chips and other equipment to process the images. In the prior art, the hough transform line detection method is used to detect the color difference R-B image of corn stalks to realize the recognition of the stalks. Although the line detection method based on Hough transform is a mature algorithm, the calculation amount is relatively large, which is not conducive to real-time The endpoint and length information of the line segment is lost during the application and transformation process.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for identifying corn stalks, which can identify corn stalks in real time and accurately, and enable the automatic fertilizer applicator to avoid equipment damage due to the obstruction of the stalks during fertilization operations Or damage to corn plants and other problems.

In order to solve the above problems, the invention provides a method for identifying corn, comprising the following steps:

(1) Collect RGB color images;

(2) converting the RGB color image into a grayscale image;

(3) Preprocessing the grayscale image;

(4) converting the preprocessed grayscale image into a binarized image through binarization processing;

(5) performing line detection on the binarized image;

(6) Calculate the shape features of the different regions of the binarized image segmented by the detected line, and identify the collected RGB color images according to the shape features of the different regions to obtain the final stalk target.

The preprocessing is wavelet lifting processing, which can extract low-frequency contour information of the gray image and suppress high-frequency noise.

When the preprocessing in step (3) is wavelet lifting processing, step (4) further includes:

Differential operator is used to binarize the grayscale image after wavelet lifting processing and pixel reduction, and the grayscale image is converted into a binary image.

Step (5) further includes:

The boundary points of the binarized image are obtained by zero-crossing detection, and then the boundary points are converted into symbol information represented by chain codes through the boundary tracking algorithm, and the line detection is performed based on the chain code string obtained by boundary tracking.

In the present invention, after the RGB color image of the corn stalk is grayscaled, wavelet lifted, and binarized, the boundary point detection is performed, and the straight line detection based on the chain code is used to identify the corn stalk in real time and accurately. Make the automatic fertilizer applicator avoid problems such as equipment damage or damage to corn plants due to the obstruction of the stalk during fertilization.

Description of drawings

Fig. 1 is the flowchart of the corn stalk identification method described in the embodiment of the present invention;

FIG. 2 is a schematic diagram of an 8-direction chain code.

Detailed ways

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

As shown in Figure 1, a kind of corn stalk identification method of the present invention comprises the following steps:

(1) Collect RGB color images;

(2) converting the RGB color image into a grayscale image;

In this step, grayscale processing is performed on the collected RGB color image, and the color image is converted into a grayscale image. Let the original color of a point in the color image be RGB(R,G,B), and it can be converted to grayscale by the following methods:

1) Floating point algorithm: Gray=R*0.3+G*0.59+B*0.11;

2) Integer method: Gray=(R*30+G*59+B*11)/100;

3) Shift method: Gray=(R*76+G*151+B*28)>>8;

4) Average method: Gray=(R+G+B)/3;

5) Take only green: Gray=G;

After obtaining Gray by any of the above methods, replace R, G, and B in the original RGB (R, G, B) with Gray to form a new color RGB (Gray, Gray, Gray), and use it to replace The original RGB (R, G, B) gets a grayscale image.

In the process of image processing, in order to use the color features of the color image to separate the target object from the image background, the color image is often subjected to dimensionality reduction processing, and only a certain feature component is used to describe the color of the image, and the color image is converted into a gray image. degree image. At the same time, the amount of image data is also significantly reduced through the grayscale of the color image, which greatly improves the image processing speed. In the present invention, since the color information of the image is not important to the morphological information of the stalk to be extracted, the grayscale image better retains the morphological information of the stalk in the original color image.

(3) Preprocessing the grayscale image;

Above-mentioned preprocessing is wavelet lifting processing, carries out wavelet lifting processing to the converted gray scale image, and described wavelet lifting processing can extract gray scale image low-frequency profile information, restrain high-frequency noise;

Wavelet lifting is a faster and more efficient way to realize wavelet transform, which is called the second generation wavelet transform. Wavelet lifting does not depend on Fourier transform, and inherits the multi-resolution characteristics of the first generation wavelet. The coefficients after wavelet transform are integers, and no additional memory is required. Operations can be performed using local operations, and wavelet transform of any image size can be realized. . All wavelet transforms can be implemented with wavelet lifting modes by means of factorized wavelet transforms.

In this step, the wavelet lifting process includes splitting, prediction and updating. Considering the signal s ^j ={s ^j,l |0≤l≤2 ^j }, the low-frequency signals obtained after one-level wavelet transformation are s ^j-1 and high-frequency signal d ^j-1 , the specific implementation process of constructing a low-resolution image is:

I． Splitting: Divide the input signal s ^j into odd and even subsets, and the set of elements at the even subscript position is recorded as even _j-1 ={s _j,2l |0≤l≤2 ^j-1 -1} , the set of elements at odd subscript positions is recorded as odd _j-1 ={s _j,2l+1 |0≤l≤2 ^j-1 -1};

II． Prediction: Based on the correlation of the original data, use the predicted value of the even sequence to predict or interpolate the odd sequence;

III. Update: The purpose of the update is to find a better subset s ^j-1 to keep a certain scalar property Q(x) of the original image (such as the root mean square value unchanged), that is, Q(s ^j-1 )=Q(s ^j );

Among them, for wavelet lifting, there are split operator Split, predictor P and update operator U, so that

(even _j-1 ，odd _j-1 )=Split(s ^j ),

d ^j-1 = odd _j-1 -P(even _j-1 ),

s ^j-1 =even _j-1 +U(d ^j-1 ).

此外，由于噪声大多分布在图像的高频部分，因此基于小波提升所构建的低分辨率图像有效地抑制了高频噪声。An image with a pixel size of M×N, after a wavelet upgrade, its low-frequency components are used to construct a low-resolution image, and its pixels are

In addition, since the noise is mostly distributed in the high-frequency part of the image, the low-resolution image constructed based on wavelet lifting effectively suppresses the high-frequency noise.

(4) converting the preprocessed grayscale image into a binarized image through binarization processing;

When the preprocessing in step (3) is wavelet lifting processing, this step further includes: using a differential operator to perform binary processing on the grayscale image after wavelet lifting processing and pixel reduction, converting the grayscale image into binarized image;

(5) performing line detection on the binarized image;

This step further includes:

The boundary points of the binarized image are obtained by zero-crossing detection, and then the boundary points are converted into symbol information represented by the chain code through the boundary tracking algorithm, and the line detection is performed based on the chain code string obtained by boundary tracking;

In this step, the boundary tracking algorithm mainly includes the scanning process and the tracking process. It is assumed that the boundary exists between pixels and is composed of edge elements. There is a horizontal edge element with (x+1, y) as the reference point, and it is judged whether there is a vertical edge element with (x, y+1) as the reference point. There is no boundary passing between them, and the value of the edge element is zero. According to whether they are target points or background points, they are divided into two types: positive zero and negative zero. The value of the edge element is non-zero, and it is positive or negative according to whether the investigation point is a target point or a background point;

The tracking direction is stipulated as follows: the target point is always on the left of the forward direction. For the outer boundary of a connected area, the tracking direction is counterclockwise, while the inner boundary (the boundary formed by the inner hole) is clockwise. When in a closed chain When calculating the area on the basis of codes, the area enclosed by the outer boundary chain codes is positive, while the product of the area enclosed by the inner boundary chain codes is negative;

The value of the edge element also indicates the corresponding tracking direction. A certain edge element direction is positive. For horizontal edge elements, the tracking direction is to the right, and for vertical edge elements, the tracking direction is upward;

The mark during the tracking process is placed on the target point of the current edge element to distinguish whether the boundary point has been tracked;

When there is a non-zero edge element between the current inspection point and the reference point, according to the situation of the remaining two points in the 2×2 window formed by these two points, the next tracking direction, that is, the subsequent edge element, is determined. When returning to the starting point , the tracking process ends.

Among them, the scanning process adopts the method of grid scanning to find the boundary points that have not been tracked, and the chain code adopts 8-direction chain code. The above-mentioned boundary tracking core algorithm greatly improves the algorithm efficiency.

When doing a grid scan of the image, only areas that intersect the grid lines are likely to be tracked, and those that fall completely within the grid are ignored. Therefore, the use of grid scanning can also greatly improve the processing speed.

In the process of straight line detection based on the chain code string obtained by boundary tracking, the following form is used to represent the chain code string of the target boundary:

I: Len,X,Y,d ₁ ,d ₂ ,…,d _n

Where I is the serial number of the current chain code string, Len is the total length of the chain code in the current chain code string, X, Y are the image coordinates of the starting point of the current chain code string, d ₁ , d ₂ ,..., d _n are the current chain code string The direction code of each pixel on the code string, the steps of straight line detection are:

IV． Minimum straight line segment length verification: sequentially judge each chain code string, if the length of the chain code string is less than the preset minimum straight line segment length threshold LT, discard the chain code string;

V． Extract a straight line segment from a chaincode string according to the straight line segment approximation criterion:

Sequentially judge each chaincode string;

①Starting from the starting point, select the sub-chain code strings that satisfy the minimum straight line segment length constraint in turn, and calculate the straight line segment approximation S of this segment according to formula (1). If S≥ST, ST is the preset minimum straight line segment similarity threshold , then the segment satisfies the linear constraint, go to ②, otherwise discard the segment and continue to judge the next sub-chain code string;

$\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; Len</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, ||p _s -p _e || represents the ideal straight-line distance between the endpoint p _s and p _e , and Len(p _s , p _e ) represents the actual pixel distance of the sub-chain code string from the endpoint p _s to the endpoint p _e length, and S represents the approximation of the straight line segment. In addition, when calculating the actual length of the chain code string, if the direction code is 0, 2, 4, or 6, the actual length between two pixels connected by the chain code is 1. If the direction The codes are 1, 3, 5, 7, then the actual length between two pixels is 2 (see Figure 2);

②If the previous sub-chain code string also satisfies the linear constraint, consider whether two adjacent sub-chain code strings can be merged into a longer straight line segment, and connect the starting point of the previous chain code string to the end point of this chain code string The chain code in between is regarded as a new chain code string, and its linear segment approximation S' is calculated. If S’≥ST, then merge, otherwise mark the sub-chain code string of this segment as a new straight line segment;

③ If the end point of the chain code string has been reached, then end, otherwise go to ① and continue to judge the next sub-chain code string.

Wherein, the preset minimum straight line segment length threshold LT is 23 pixels, and the preset minimum straight line segment similarity threshold ST is 0.91.

The present invention first traces the chain code on the boundary of the target, and then extracts the straight line segment from the obtained chain code string set. The straight line detection algorithm adopted in the present invention starts from practical application, gets rid of the constraint of Freeman's ideal straight line criterion, only adopts two constraint parameters of minimum straight line segment length and minimum straight line similarity, and can quickly and accurately realize straight line detection. This algorithm has fast detection speed and strong practicability, which enhances the real-time performance of the corn stalk identification process.

(6) Calculate the shape features of the different areas of the binarized image segmented by the detected line, and identify the collected RGB color images according to the shape features of the different areas to obtain the final stalk target.

和偏心率进行识别，满足F大于2.3且e大于5的区域识别为茎秆的步骤，其中，Area和Per分别为区域的面积和周长，c和a分别为区域的边界长轴长度和短轴长度，长轴是指区域边界上距离最远的两点连线，区域边界上与长轴垂直的连线中最长的线段称为短轴。In this step, the shape feature has multiple parameters, and in this embodiment, only shape parameters and eccentricity are calculated. Identify the area where the line is located as a stalk, and use the shape parameters for other areas divided by the line

and eccentricity Identify the step of identifying the area that satisfies F greater than 2.3 and e greater than 5 as a stalk, where Area and Per are the area and perimeter of the area respectively, c and a are the length of the major axis and the length of the minor axis of the boundary of the area, respectively, The long axis refers to the line connecting the two points farthest on the boundary of the region, and the longest line segment among the lines perpendicular to the long axis on the boundary of the region is called the short axis.

The above embodiments are only used to illustrate the present invention, but not to limit the present invention. Those of ordinary skill in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all Equivalent technical solutions also belong to the category of the present invention, and the scope of patent protection of the present invention should be defined by the claims.

Claims (10)

1. a method for automatically identifying corn stem and leaf images, is characterized in that, comprises the following steps: (1) Collect RGB color images; (2) converting the RGB color image into a grayscale image; (3) Preprocessing the grayscale image; (4) converting the preprocessed grayscale image into a binarized image through binarization processing; (5) performing line detection on the binarized image; (6) Calculate the shape features of the different regions of the binarized image segmented by the detected line, and identify the collected RGB color images according to the shape features of the different regions to obtain the final stalk target. 2. The method according to claim 1, wherein the preprocessing is wavelet lifting processing, which can extract low-frequency contour information of grayscale images and suppress high-frequency noise. 3. The method according to claim 2, the wavelet lifting process includes splitting, prediction and updating, considering the signal s ^j ={s _{j, l} |0≤l≤2 ^j }, which is obtained after one-stage wavelet transformation The low-frequency signal is s ^j-1 and the high-frequency signal d ^j-1 , then the specific implementation process of constructing a low-resolution image is: I． Splitting: Divide the input signal s ^j into odd and even subsets, and the set of elements at the even subscript position is recorded as even _j-1 ={s _j,2l |0≤l≤2 ^j-1 -1} , the set of elements at odd subscript positions is recorded as odd _j-1 ={s _j,2l+1 |0≤l≤2 ^j-1 -1}; II． Prediction: Based on the correlation of the original data, use the predicted value of the even sequence to predict or interpolate the odd sequence; III. Update: The purpose of the update is to find a better subset s ^j-1 to keep a certain scalar property Q(x) of the original image (such as the root mean square value unchanged), that is, Q(s ^j-1 )=Q(s ^j ); Among them, for wavelet lifting, there are split operator Split, predictor P and update operator U, so that (even _j-1 ，odd _j-1 )=Split(s ^j ), d ^j-1 = odd _j-1 -P(even _j-1 ), s ^j-1 =even _j-1 +U(d ^j-1 ). 4. The method as claimed in claim 3, wherein said keeping a certain scalar characteristic Q(x) of the original image as a constant root mean square value. 5. The method according to claim 2, 3 or 4, step (4) further comprising: Differential operator is used to binarize the grayscale image after wavelet lifting processing and pixel reduction, and the grayscale image is converted into a binary image. 6. The method according to claim 1, step (5) further comprising: The boundary points of the binarized image are obtained by zero-crossing detection, and then the boundary points are converted into symbol information represented by chain codes through the boundary tracking algorithm, and the line detection is performed based on the chain code string obtained by boundary tracking. 7. The method as claimed in claim 6, the boundary tracking algorithm mainly includes a scanning process and a tracking process, assumes that the boundary exists between pixels, is composed of edge elements, and is divided into horizontal and vertical two kinds, assuming that the current investigation point coordinates is (x, y), judge whether there is a horizontal edge element with (x+1, y) as the reference point, and judge whether there is a vertical edge element with (x, y+1) as the reference point, when the distance between the inspection point and the reference point When the values are the same, there is no boundary between the two points, and the value of the edge element is zero. According to whether they are target points or background points, they are divided into positive zero and negative zero. When the values of the inspection point and the reference point are different When there is a boundary passing between the two points, the value of the edge element is non-zero, and it is positive or negative according to whether the investigation point is a target point or a background point; The tracking direction is stipulated as follows: the target point is always on the left side of the forward direction. For the outer boundary of a connected area, the tracking direction is counterclockwise, while the inner boundary is clockwise. When calculating the area on the basis of closed chain codes, the outer boundary The area enclosed by the boundary chain codes is positive, while the product of the area enclosed by the inner boundary chain codes is negative; The value of the edge element also indicates the corresponding tracking direction. A certain edge element direction is positive. For horizontal edge elements, the tracking direction is to the right, and for vertical edge elements, the tracking direction is upward; The mark during the tracking process is placed on the target point of the current edge element to distinguish whether the boundary point has been tracked; When there is a non-zero edge element between the current inspection point and the reference point, according to the situation of the remaining two points in the 2×2 window formed by these two points, determine the next tracking direction, that is, the subsequent edge element, when returning to the starting point , the tracking process ends. 8. The method according to claim 7, wherein the scanning process adopts grid scanning to find boundary points not yet tracked, the chain code is an 8-direction chain code, and the inner boundary is a boundary formed by an internal hole . 9. The method according to claim 6, in the chain code string obtained based on boundary tracking, in the straight line detection process, the chain code string of the target boundary is represented in the following form: I: Len,X,Y,d ₁ ,d ₂ ,…,d _n Where I is the serial number of the current chain code string, Len is the total length of the chain code in the current chain code string, X, Y are the image coordinates of the starting point of the current chain code string, d ₁ , d ₂ ,..., d _n are the current chain code string The direction code of each pixel on the code string, the steps of straight line detection are: IV． Minimum straight line segment length verification: sequentially judge each chain code string, if the length of the chain code string is less than the preset minimum straight line segment length threshold LT, discard the chain code string; V． Extract a straight line segment from a chaincode string according to the straight line segment approximation criterion: Sequentially judge each chaincode string; ①Starting from the starting point, select the sub-chain code strings that satisfy the minimum straight line segment length constraint in turn, and calculate the straight line segment approximation S of this segment according to formula (1). If S≥ST, ST is the preset minimum straight line segment similarity threshold , then the segment satisfies the linear constraint, go to ②, otherwise discard the segment and continue to judge the next sub-chain code string; $\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; Len</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$ Among them, ||p _s -p _e || represents the ideal straight-line distance between the endpoint p _s and p _e , and Len(p _s , p _e ) represents the actual pixel distance of the sub-chain code string from the endpoint p _s to the endpoint p _e length, and S represents the approximation of the straight line segment. In addition, when calculating the actual length of the chain code string, if the direction code is 0, 2, 4, or 6, the actual length between two pixels connected by the chain code is 1. If the direction The code is 1, 3, 5, 7, then the actual length between two pixels is 2; ②If the previous sub-chain code string also satisfies the linear constraint, consider whether two adjacent sub-chain code strings can be merged into a longer straight line segment, and connect the starting point of the previous chain code string to the end point of this chain code string The chain code in between is regarded as a new chain code string, and its linear segment approximation S' is calculated. If S’≥ST, then merge, otherwise mark the sub-chain code string of this segment as a new straight line segment; ③ If it has reached the end of the chain code string, it will end, otherwise go to ① to continue to judge the next sub-chain code string; Wherein the preset minimum straight line segment length threshold LT is 23 pixels, and the preset minimum straight line segment similarity threshold ST is 0.91. 10. The method according to claim 1, said step (6) further comprising: identifying the area where the straight line is located as a stalk, and using shape parameters for other areas divided by the straight line and eccentricity

The step of identifying the region that satisfies F greater than 2.3 and e greater than 5 is identified as a stalk, where Area and Per are the area and perimeter of the region, and c and a are the length of the major axis and minor axis of the region's boundary, respectively.

Images