Sample point scale cotton field boll opening rate monitoring method

Abstract

The invention discloses a sampling point scale cotton field boll opening rate monitoring method, and relates to the field of cotton field boll opening rate monitoring. The method comprises the following steps: acquiring images before and after a ripener is sprayed to a cotton field, and acquiring a boll opening rate in the same period; segmenting the image into sub-images and marking the sub-images, performing cotton boll identification on the sub-images by using various models, and determining an optimal model; re-segmenting the image into sub-images with multiple pixel sizes, identifying and labeling the sub-images with all pixel sizes by adopting an optimal model, recombining the sub-images into an original image, and keeping a cotton boll labeling frame; grouping the recombined original images according to the pixel size, selecting the images from the groups, combining and overlapping the images in pairs, forming a new cotton boll marking frame by judging the cotton boll marking frame, and calculating the boll opening rate of the overlapped images; and comparing fitting effects of boll opening rates of all overlapped images, obtaining an optimal pixel size combination, and obtaining an optimal shooting scheme by adopting the optimal pixel size combination, thereby realizing accurate and rapid monitoring of the boll opening rate of the cotton field.

Description

Technical Field

[0001] The invention relates to the field of agriculture, and in particular to a cotton field boll opening rate monitoring method at a sampling point scale. Background Art

[0002] Cotton is an important economic crop, widely planted around the world. The application of harvesting aids for defoliation and ripening is a prerequisite for mechanical harvesting of cotton, and the determination of the spraying time and dosage of harvesting aids requires comprehensive consideration of multiple factors. Among them, the boll opening rate is a key factor in determining the timing and dosage of cotton defoliation and ripening agents, as well as evaluating the ripening effect and determining the harvesting time. At present, the boll opening rate is required to reach 30-40% before spraying defoliation and ripening agents in cotton producing areas in China. In addition, the boll opening rate is also an important indicator for evaluating the application effect of harvesting aids and judging the time of mechanical harvesting. The traditional method of manually investigating the boll opening rate is to first select sample points, determine the survey plants, and then count the boll opening bolls and non-boll opening bolls of each plant, and then calculate the boll opening rate by dividing the number of boll opening bolls by the total number of bolls (the sum of the boll opening bolls and the non-boll opening rate). The traditional method is labor-intensive, time-consuming, and inefficient. The number of sample points surveyed per unit time is limited, so the representativeness is also poor. Therefore, it is urgent to develop a method for quickly and accurately monitoring the boll opening rate of cotton fields.

[0003] In the existing technology, with the upgrade of deep learning algorithms and the improvement of computing power of personal computing devices, the research on the boll opening rate is mainly based on image recognition and the construction of 3D point clouds of cotton fields to achieve the segmentation and counting of cotton bolls. Based on YOLOX, researchers introduced multi-scale residual modules and attention modules to enhance the extraction of cotton field image feature details. In order to reduce the recognition loss of small targets, multi-receptive field extraction modules were added to realize the image recognition of un-bolled cotton bolls. Some researchers used self-created drone routes to obtain RGB images of cotton fields, and built 3D point clouds of cotton fields based on them. They also used histograms of RGB and HSV color space channels and cotton boll shapes, spatial positions, etc. for feature extraction, and used deep forest classification to finally realize the recognition of bolled and un-bolled cotton bolls.

[0004] However, the above-mentioned prior art requires compression or segmentation when processing images. Compression will cause the loss of detailed features of cotton bolls, and segmentation will cause misidentification or missed recognition of border bolls. In addition, cotton fields have serious occlusion in the late growth period. The prior art uses drones to obtain frontal images of cotton bolls. It is difficult to solve this problem. There are a large number of missed recognition problems in the lower and middle parts of the canopy in the images.

[0005] In summary, the existing technology has the problems of loss of detailed features of cotton bolls, misidentification and missed identification, which leads to reduced recognition ability of cotton bolls and makes the monitoring of cotton field boll opening rate not accurate and fast enough. Summary of the invention

[0006] Based on this, it is necessary to provide a cotton field boll opening rate monitoring method at a sampling point scale to address the above technical issues.

[0007] The present invention adopts the following technical solutions:

[0008] A cotton field boll opening rate monitoring method comprises the following steps:

[0009] Obtain multiple original images of cotton fields before and after the ripening agent is sprayed, and obtain the true value of the cotton field's boll opening rate during the same period;

[0010] The original image is divided into a first sub-image according to a specified pixel size and the cotton bolls are marked for the first time; a plurality of deep learning models for target detection are used to identify the cotton bolls in the first sub-image to obtain the identification results; and the optimal deep learning model is determined according to the identification results;

[0011] The original image is re-divided into second sub-images of multiple pixel sizes, and the cotton bolls are identified and labeled for the second sub-images using the optimal deep learning model; the second sub-images of each pixel size are reassembled by splicing to obtain a reassembled original image of each pixel size; wherein the reassembled original image of each pixel size retains the cotton boll labeling box of its second sub-image;

[0012] The reconstructed original images of all pixel sizes are divided into two groups according to the pixel size, and the reconstructed original images of one pixel size are selected from each of the two groups for pairwise combination and overlap, so as to obtain an overlapped image of each pixel size combination;

[0013] Determine the overlap of the cotton boll annotation frames of the two reconstructed original images in the overlapping images of each pixel size combination; according to the determination result, replace the cotton boll annotation frame in the overlapping image with a new cotton boll annotation frame, obtain each overlapping image with the new cotton boll annotation frame, and calculate the cotton boll opening rate of each overlapping image through the cotton bolls in the new cotton boll annotation frame;

[0014] By comparing the fitting effect of the boll opening rate of each overlapping image with the true value of the boll opening rate of the cotton field in the same period, the optimal pixel size combination of the two reconstructed original images in the overlapping image is obtained; the boll opening rate of the cotton field is monitored through the obtained optimal pixel size combination.

[0015] Preferably, when the original images of the cotton field before and after the ripening agent is sprayed are obtained by ground photography, the method further comprises:

[0016] The best shooting scheme parameters for the original image are obtained by comparing the fitting effect of the overlapping image boll opening rate of the best pixel size combination under different shooting scheme parameters with the real value of the boll opening rate of the cotton field in the same period.

[0017] The cotton field boll opening rate is monitored by obtaining the optimal pixel size combination and the optimal shooting plan parameters.

[0018] Preferably, the shooting scheme parameters use the top of the cotton field canopy as a reference point, and the shooting scheme parameters include the height, angle and shooting direction set during shooting.

[0019] Preferably, the investigation of the true value of the ground fluffing rate of each sampling point during the same period specifically includes:

[0020] Cotton plants with uniform growth were selected at each sampling point. After the shooting was completed, the total number of bolls and the number of bolls opened of each plant were counted. The true value of the ground boll opening rate was the ratio of the number of bolls opened to the total number of bolls.

[0021] Preferably, obtaining the overlapping images of each pixel size combination comprises the following steps:

[0022] Using the middle values ​​of the plurality of pixel sizes as a dividing line, the reconstructed original images of all pixel sizes are divided into a first group and a second group;

[0023] Selecting a recombined original image of a pixel size in each of the first group and the second group, respectively, and combining and overlapping them in pairs, to obtain an overlapping image of a pixel size combination;

[0024] Repeatedly selecting recombined original images from the two groups to combine and overlap them in pairs until an overlapping image of each pixel size combination is obtained; wherein the pixel size combinations of the two recombined original images in each overlapping image are different.

[0025] Preferably, the step of acquiring all overlapping images with new cotton boll annotation frames specifically includes:

[0026] For each overlapping image of each pixel size combination, the overlapping of the cotton boll annotation frames of the recombined original images of two pixel sizes in the overlapping image is determined to form a new cotton boll annotation frame of the overlapping image, including:

[0027] If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the first group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the second group, then the cotton boll labeling frame of the recombined original image of one pixel size in the first group is retained to form a new cotton boll labeling frame of the overlapping image;

[0028] If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the second group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the first group, then the cotton boll labeling frame of the recombined original image of one pixel size in the second group is retained to form a new cotton boll labeling frame of the overlapping image;

[0029] If there is an overlap between the cotton boll annotation frames of the two recombined original images in the overlapped image, the outermost circle positions of the cotton boll annotation frames of the two recombined original images are calculated and the cotton bolls are annotated to form a new cotton boll annotation frame of the overlapped image;

[0030] The cotton boll labeling frame of the reconstructed original image in the overlapped image is replaced by the new cotton boll labeling frame of the overlapped image formed in the above manner, so as to obtain the overlapped image with the new cotton boll labeling frame.

[0031] Preferably, the formation of the new cotton boll annotation frame of the overlapping image specifically includes:

[0032] Taking the lower left corner of the image as the origin, vertically upward as the positive direction of the Y axis, and horizontally to the right as the positive direction of the X axis, the coordinates of the four corner points of the cotton boll annotation box of the reorganized original image of one pixel size in the first group are the lower left corner (xa1, ya1), the upper left corner (xa2, ya2), the upper right corner (xa3, ya3) and the lower right corner (xa4, ya4); the cotton boll annotation box of the reorganized original image of one pixel size in the second group adopts the same order, and the coordinates of its four corner points are recorded as (xb1, yb1), (xb2, yb2), (xb3, yb3) and (xb4, yb4);

[0033] Compare the sizes of the 8 groups of values ​​in the 4 groups of coordinates respectively, retain the smaller values ​​in xa1 and xb1, ya1 and yb1, xa2 and xb2, and ya4 and yb4, and record them as xc1, yc1, xc2, and yc4 respectively; retain the larger values ​​in ya2 and yb2, xa3 and xb3, ya3 and yb3, and xa4 and xb4, and record them as yc2, xc3, yc3, and xc4 respectively;

[0034] Through the above steps, four new sets of coordinates are obtained, including: lower left corner coordinates (xc1, yc1), upper left corner coordinates (xc2, yc2), upper right corner coordinates (xc3, yc3), and lower right corner coordinates (xc4, yc4);

[0035] The quadrilateral formed by the four sets of new coordinates is the new cotton boll marking frame of the overlapping image.

[0036] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0037] The present invention divides the original image into sub-images of multiple pixel sizes, performs model recognition and annotation on the sub-images and reorganizes them into original images of various pixel sizes, and groups, combines and overlaps the original images according to the pixel sizes to obtain overlapping images of different pixel combinations. By judging the sub-image cotton boll annotation frames carried by the overlapping images, the incomplete cotton bolls at the edge of the image can be spliced ​​into a complete cotton boll, wherein the larger sub-image cotton boll annotation frame included in the overlapping image may cover the entire cotton boll, and the smaller sub-image cotton boll annotation frame can be used to capture the image detail features of the small cotton boll, thereby solving the problem of loss of cotton boll detail features and wrong recognition and missed recognition of cotton boll boundaries caused by image compression and segmentation.

[0038] In addition, the present invention adopts the method of ground shooting to more easily obtain image information of the middle and lower parts of the cotton boll canopy, making the obtained image data information more complete, thereby avoiding the occlusion phenomenon existing in the cotton field in the late growth period.

[0039] In summary, the present invention improves the recognition ability of cotton boll identification by combining and overlapping cotton boll images with different pixel sizes to obtain the optimal pixel size combination and determine the optimal shooting plan, thereby achieving accurate and rapid monitoring of the cotton field boll opening rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 A schematic flow chart of a cotton field boll opening rate monitoring method at a sampling point scale provided by the present invention;

[0042] Figure 2 A schematic diagram of an image shooting scheme for a cotton field boll opening rate monitoring method at a sample point scale provided by the present invention;

[0043] Figure 3 The recognition results of the YOLOv5m model of the cotton field boll opening rate monitoring method at a sample point scale provided by the present invention on sub-images of all pixel sizes;

[0044] Figure 4 The comparison of recognition effects before and after the merging operation of 400 and 700 pixel sub-images of a cotton field boll opening rate monitoring method at a sample point scale provided by the present invention;

[0045] Figure 5 The present invention provides a cotton field boll opening rate monitoring method at a sample point scale, and provides a fitting effect of the boll opening rate of overlapping images of all pixel size combinations with the true value. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0047] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0048] Figure 1 The present invention is a schematic flow chart of a cotton field boll opening rate monitoring method at a sampling point scale, which specifically includes the following steps:

[0049] S101: obtaining multiple original images of the cotton field before and after the ripening agent is sprayed, and obtaining the true value of the cotton field boll opening rate during the same period;

[0050] Among them, the acquisition of the true value of the cotton field boll opening rate during the same period includes: selecting 20 cotton plants with uniform growth at each sampling point, and investigating the total number of bolls and the number of bolls opening for each plant after shooting. The true value of the ground boll opening rate is the ratio of the number of bolls opening to the total number of bolls.

[0051] S102: Divide the original image into a first sub-image according to a specified pixel size and perform a first cotton boll labeling; use a variety of deep learning models for target detection to perform cotton boll recognition on the first sub-image to obtain recognition results; determine the optimal deep learning model based on the recognition results.

[0052] S103: Re-segmenting the original image into a plurality of second sub-images of pixel sizes, using an optimal deep learning model to perform cotton boll recognition and second cotton boll labeling on the second sub-images, and recombining the second sub-images of each pixel size into the original image by splicing, to obtain a recombined original image of each pixel size; wherein the recombined original image of each pixel size retains the cotton boll labeling box of its second sub-image;

[0053] S104: Divide the reconstructed original images of all pixel sizes into two groups, select one reconstructed original image of each pixel size from each group, combine them in pairs and overlap them, and obtain overlapping images of each pixel size combination, specifically including:

[0054] Using the middle values ​​of the plurality of pixel sizes as a dividing line, the reconstructed original images of all pixel sizes are divided into a first group and a second group;

[0055] Selecting a recombined original image of a pixel size in each of the first group and the second group, respectively, and combining and overlapping them in pairs, to obtain an overlapping image of a pixel size combination;

[0056] Repeatedly selecting recombined original images from the two groups to combine and overlap them in pairs until an overlapping image of each pixel size combination is obtained; wherein the pixel size combinations of the two recombined original images in each overlapping image are different.

[0057] S105: determining the overlap of the cotton boll annotation frames of the two reconstructed original images in the overlapping images of each pixel size combination, and according to the determination result, replacing the cotton boll annotation frame in the overlapping image with a new cotton boll annotation frame, obtaining each overlapping image with the new cotton boll annotation frame, and calculating the cotton boll opening rate of each overlapping image through the cotton bolls in the new cotton boll annotation frame, specifically including:

[0058] For each overlapping image of each pixel size combination, the overlapping of the cotton boll annotation frames of the recombined original images of two pixel sizes in the overlapping image is determined to form a new cotton boll annotation frame of the overlapping image, including:

[0059] For each overlapping image of each pixel size combination, the overlapping of the cotton boll annotation frames of the recombined original images of two pixel sizes in the overlapping image is determined to form a new cotton boll annotation frame of the overlapping image, including:

[0060] If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the first group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the second group, then the cotton boll labeling frame of the recombined original image of one pixel size in the first group is retained to form a new cotton boll labeling frame of the overlapping image;

[0061] If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the second group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the first group, then the cotton boll labeling frame of the recombined original image of one pixel size in the second group is retained to form a new cotton boll labeling frame of the overlapping image;

[0062] If there is an overlap between the cotton boll annotation frames of the two recombined original images in the overlapped image, the outermost circle positions of the cotton boll annotation frames of the two recombined original images are calculated and the cotton bolls are annotated to form a new cotton boll annotation frame of the overlapped image;

[0063] The new cotton boll annotation frame of the overlapped image formed above replaces the cotton boll annotation frame of the reconstructed original image in the overlapped image to obtain an overlapped image with the new cotton boll annotation frame;

[0064] The formation of a new cotton boll annotation frame of the overlapping image includes:

[0065] Taking the lower left corner of the image as the origin, vertically upward as the positive direction of the Y axis, and horizontally to the right as the positive direction of the X axis, the coordinates of the four corner points of the cotton boll annotation box of the reorganized original image of one pixel size in the first group are the lower left corner (xa1, ya1), the upper left corner (xa2, ya2), the upper right corner (xa3, ya3) and the lower right corner (xa4, ya4); the cotton boll annotation box of the reorganized original image of one pixel size in the second group adopts the same order, and the coordinates of its four corner points are recorded as (xb1, yb1), (xb2, yb2), (xb3, yb3) and (xb4, yb4);

[0066] Compare the sizes of the 8 groups of values ​​in the 4 groups of coordinates respectively, retain the smaller values ​​in xa1 and xb1, ya1 and yb1, xa2 and xb2, and ya4 and yb4, and record them as xc1, yc1, xc2, and yc4 respectively; retain the larger values ​​in ya2 and yb2, xa3 and xb3, ya3 and yb3, and xa4 and xb4, and record them as yc2, xc3, yc3, and xc4 respectively;

[0067] Through the above steps, four new sets of coordinates are obtained, including: lower left corner coordinates (xc1, yc1), upper left corner coordinates (xc2, yc2), upper right corner coordinates (xc3, yc3), and lower right corner coordinates (xc4, yc4);

[0068] The quadrilateral formed by the four sets of new coordinates is the new cotton boll marking frame of the overlapping image.

[0069] S106: Obtain the best pixel size combination of two reconstructed original images in the overlapping images by comparing the fitting effect of the boll opening rate of each overlapping image with the true value of the boll opening rate of the cotton field in the same period; and monitor the boll opening rate of the cotton field through the obtained best pixel size combination.

[0070] S107: Obtain the best shooting scheme parameters of the original image by comparing the fitting effect of the cotton-opening rate of the overlapping images of the best pixel size combination under different shooting scheme parameters with the real cotton-opening rate of the cotton field during the same period; and monitor the cotton-opening rate of the cotton field by using the obtained best pixel size combination and the best shooting scheme parameters;

[0071] Among them, the shooting scheme parameters take the top of the cotton field canopy as the reference point, and the parameters set during shooting include height, angle and shooting direction;

[0072] In addition, the shooting scheme of the above cotton field original image is as follows: Figure 2 As shown, it includes: setting 4 heights, 5 angles and 2 shooting directions above the cotton field canopy to shoot the cotton field, and obtaining original images of the cotton field at multiple periods before and after the ripening agent is sprayed;

[0073] Among them, there are 4 heights, including: 0cm, 10cm, 20cm and 30cm; 5 angles, including: 0°, 15°, 30°, 45° and 60°; 2 shooting directions, including: parallel to the planting rows and perpendicular to the planting rows; among them, the direction of the reference angle 0° is horizontal to the right, and the 5 angles are set in clockwise order according to the direction of the reference angle.

[0074] Taking the Chinese cotton H318 in a certain place in 2022 as an example, it was sown on April 25 and field management was carried out according to local practices. On September 23, the harvesting auxiliary agent 50% thiophene·ethephon suspension (T·E) was sprayed, with four treatments of 750, 1500, 3000, and 4500 mL·hm-2, and clear water was used as the control.

[0075] Images were collected on September 15, September 30, October 7, and October 14, and the fluffing rate was manually investigated. The camera used was Canon eos550 (P mode, aperture value f / 13, exposure time 1 / 160s, focal length 16mm, ISO-400). After selecting the sample points, photos were taken at different heights above the canopy using different angles and shooting directions, and 912 photos with a resolution of 5184*3456 pixels were obtained. The shooting height was set to 4 levels of 0, 10, 20, and 30cm, and the shooting angle (angle downward with the horizontal plane) was set to 5 levels of 0°, 15°, 30°, 45°, and 60°. The shooting direction was divided into 2 types: parallel to the planting row and perpendicular to the planting row. When shooting, avoid large-area occlusion of a single leaf in front of the lens;

[0076] According to the method in step S102, the model with the best comprehensive performance in recognition accuracy and recognition time is selected as the deep learning model for optimal target detection. The recognition results are shown in the following table:

[0077]

[0078] Using 8121 images (500*500 pixels) as training data sets, the recognition accuracy of the target detection model YOLO v3, the fast regional convolutional neural network model Faster-RCNN, the target detection model YOLOv5 and Centernet all reached an acceptable level (precision>0.8). Among them, the target detection model YOLOv5 and the fast regional convolutional neural network model Faster-RCNN have higher recognition accuracy, and their classification accuracy and mAP_0.5 are close to 0.9. In terms of comprehensive recognition time, YOLOv5 can obtain satisfactory recognition accuracy at a relatively fast speed. Faster-RCNN, the fast regional convolutional neural network model, although Faster-RCNN ranks first in all accuracy evaluation indicators, but its recognition time is 3-5 times that of the target detection model YOLOv5. Taking all factors into consideration, the target detection model YOLOv5 is the best deep learning model for target detection;

[0079] By the method in step S103 to step S105, an overlapping image of each pixel size combination is obtained;

[0080] Among them, in step S103, several pixel sizes used in this embodiment include: 100*100, 200*200, 300*300, 400*400, 500*500, 600*600, 700*700, 800*800, 900*900 and 1000*1000;

[0081] In addition, in step S104, the reorganized original images with pixel sizes of 100*100, 200*200, 300*300, 400*400 and 500*500 are divided into the first group, and the reorganized original images with other pixel sizes are divided into the second group, taking the reorganized original image with pixel size of 500*500 as the dividing line.

[0082] The fitting effect of the calculated linting rate of the overlapping images of each pixel size combination and the true value is shown in the figure Figure 3 As shown in Figure 2, when the sub-image size is between 100 and 200 pixels, the R of the fitting effect is the same regardless of the pixel size combination of the image. 2 The values ​​were relatively low (46%) <R 2 <72%). This indicates that when the sub-image size is small, the recognition accuracy may be low due to the lack of detail information or the limitation of the recognition algorithm, which in turn affects the overall fitting effect. However, in the combination of sub-image sizes of 300 to 400 pixels and 600 to 700 pixels, and 500 pixels and 700 to 900 pixels, it can be seen that the fitting effect is greatly improved, R 2The values ​​are generally high, and the best effect is the overlapping image of 400*400 and 700*700 pixels. 2 It reached 82.80%. The recognition effect comparison before and after the combination of 400*400 and 700*700 pixels is as follows: Figure 4 shown.

[0083] The cotton bolls were identified by combining 400 and 700 pixel sizes. The cotton boll opening rates of various shooting schemes were calculated and fitted with the true values ​​to determine the influence of the shooting scheme parameters (shooting height, angle, direction) on the cotton boll opening rate fitting effect. The fitting results are shown in Figure 2. Figure 5 As shown. Among them, the closer the data point is to the upper left corner, the better the fitting effect. In all data sets and data sets shot perpendicular to the row direction, the RRMSE of the fitting model with a shooting angle of 0-30° (the angle with the horizontal direction) is all concentrated on the left side of the X-axis, while 45° and 60° are more loosely distributed on the right side; the RRMSE of the fitting model shot at a distance of 20 and 30 cm from the canopy is concentrated on the left as a whole, and some combinations of 0 and 10 cm are also concentrated on the left. Compared with shooting perpendicular to the row direction, the combination of shooting parallel to the row direction (the combination of shooting angle and height) has a higher overall recognition accuracy, among which the RRMSE of the 8 combinations is higher than that of the 10 combinations. 2 It reaches above 0.9, and the RRMSE of the seven combinations is less than 10. Specifically, the parameters with better recognition effects are the combination of 30 cm and 0, 15, 30°, and the combination of 20 cm and 0, 15, 30°.

[0084] In summary, the present invention provides a cotton field boll opening rate monitoring method at a sample point scale, which divides the original image into sub-images of multiple pixel sizes, performs model recognition and annotation on the sub-images and reorganizes them into original images of each pixel size, and groups, combines and overlaps the original images according to the pixel size to obtain overlapping images of different pixel combinations, and by judging the sub-image cotton boll annotation frame carried by the overlapping image, the incomplete cotton bolls at the edge of the image can be spliced ​​into a complete cotton boll, wherein the larger sub-image cotton boll annotation frame included in the overlapping image may cover the entire cotton boll, and the smaller sub-image cotton boll annotation frame can be used to capture the image details of the small cotton boll, thereby solving the problem of cotton boll boundary misidentification and missed identification caused by image segmentation. In addition, the use of ground shooting can make it easier to obtain image information of the middle and lower parts of the cotton boll canopy, so that the obtained image data information is more complete, thereby avoiding the occlusion phenomenon existing in the cotton field in the late growth period. In summary, the present invention improves the recognition ability of cotton boll identification by combining and overlapping cotton boll images with different pixel sizes to obtain the optimal pixel size combination and determining the optimal shooting plan, thereby realizing accurate and rapid monitoring of the cotton field boll opening rate, thereby providing powerful scientific guidance for cotton field boll opening rate monitoring.

[0085] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A cotton field boll opening rate monitoring method, characterized in that: The method comprises the following steps: Obtain multiple original images of cotton fields before and after the ripening agent is sprayed, and obtain the true value of the cotton field's boll opening rate during the same period; The original image is divided into a first sub-image according to a specified pixel size and the cotton bolls are marked for the first time; a plurality of deep learning models for target detection are used to identify the cotton bolls in the first sub-image to obtain the identification results; and the optimal deep learning model is determined according to the identification results; The original image is re-divided into second sub-images of multiple pixel sizes, and the cotton bolls are identified and labeled for the second sub-images using the optimal deep learning model; the second sub-images of each pixel size are reassembled by splicing to obtain a reassembled original image of each pixel size; wherein the reassembled original image of each pixel size retains the cotton boll labeling box of its second sub-image; The reconstructed original images of all pixel sizes are divided into two groups according to the pixel size, and the reconstructed original images of one pixel size are selected from each of the two groups for pairwise combination and overlap, so as to obtain an overlapped image of each pixel size combination; Determine the overlap of the cotton boll annotation frames of the two reconstructed original images in the overlapping images of each pixel size combination; according to the determination result, replace the cotton boll annotation frame in the overlapping image with a new cotton boll annotation frame, obtain each overlapping image with the new cotton boll annotation frame, and calculate the cotton boll opening rate of each overlapping image through the cotton bolls in the new cotton boll annotation frame; By comparing the fitting effect of the boll opening rate of each overlapping image with the true value of the boll opening rate of the cotton field in the same period, the optimal pixel size combination of the two reconstructed original images in the overlapping image is obtained; the boll opening rate of the cotton field is monitored through the obtained optimal pixel size combination.

2. A cotton field boll opening rate monitoring method as claimed in claim 1, characterized in that: When the original images of the cotton field before and after the ripening agent is sprayed are obtained by ground photography, the method further includes: The best shooting scheme parameters for the original image are obtained by comparing the fitting effect of the overlapping image boll opening rate of the best pixel size combination under different shooting scheme parameters with the real value of the boll opening rate of the cotton field in the same period. The cotton field boll opening rate is monitored by obtaining the optimal pixel size combination and the optimal shooting plan parameters.

3. A cotton field boll opening rate monitoring method as claimed in claim 2, characterized in that: The shooting scheme parameters take the top of the cotton field canopy as a reference point, and the shooting scheme parameters include the height, angle and shooting direction set during shooting.

4. A cotton field boll opening rate monitoring method as claimed in claim 1, characterized in that: The actual value of the ground fluffing rate at each sampling point during the same period of the survey specifically includes: Cotton plants with uniform growth were selected at each sampling point. After the shooting was completed, the total number of bolls and the number of bolls opened of each plant were counted. The true value of the ground boll opening rate was the ratio of the number of bolls opened to the total number of bolls.

5. A cotton field boll opening rate monitoring method as claimed in claim 1, characterized in that: The acquisition of the overlapping images of each pixel size combination comprises the following steps: Using the middle values ​​of the plurality of pixel sizes as a dividing line, the reconstructed original images of all pixel sizes are divided into a first group and a second group; Selecting a recombined original image of a pixel size in each of the first group and the second group, respectively, and combining and overlapping them in pairs, to obtain an overlapping image of a pixel size combination; Repeatedly selecting recombined original images from the two groups to combine and overlap them in pairs until an overlapping image of each pixel size combination is obtained; wherein the pixel size combinations of the two recombined original images in each overlapping image are different.

6. A cotton field boll opening rate monitoring method as claimed in claim 1, characterized in that: The step of obtaining all overlapping images with new cotton boll annotation frames specifically includes: For each overlapping image of each pixel size combination, the overlapping of the cotton boll annotation frames of the recombined original images of two pixel sizes in the overlapping image is determined to form a new cotton boll annotation frame of the overlapping image, including: If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the first group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the second group, then the cotton boll labeling frame of the recombined original image of one pixel size in the first group is retained to form a new cotton boll labeling frame of the overlapping image; If in the overlapping image, the cotton boll labeling frame of the recombined original image of one pixel size in the second group does not overlap with the cotton boll labeling frame of the recombined original image of any pixel size in the first group, then the cotton boll labeling frame of the recombined original image of one pixel size in the second group is retained to form a new cotton boll labeling frame of the overlapping image; If there is an overlap between the cotton boll annotation frames of the two recombined original images in the overlapped image, the outermost circle positions of the cotton boll annotation frames of the two recombined original images are calculated and the cotton bolls are annotated to form a new cotton boll annotation frame of the overlapped image; The cotton boll labeling frame of the reconstructed original image in the overlapped image is replaced by the new cotton boll labeling frame of the overlapped image formed in the above manner, so as to obtain the overlapped image with the new cotton boll labeling frame.

7. A cotton field boll opening rate monitoring method as claimed in claim 1 or 6, characterized in that: The formation of the new cotton boll annotation frame of the overlapping image specifically includes: Taking the lower left corner of the image as the origin, vertically upward as the positive direction of the Y axis, and horizontally to the right as the positive direction of the X axis, the coordinates of the four corner points of the cotton boll annotation box of the reorganized original image of one pixel size in the first group are the lower left corner (xa1, ya1), the upper left corner (xa2, ya2), the upper right corner (xa3, ya3) and the lower right corner (xa4, ya4); the cotton boll annotation box of the reorganized original image of one pixel size in the second group adopts the same order, and the coordinates of its four corner points are recorded as (xb1, yb1), (xb2, yb2), (xb3, yb3) and (xb4, yb4); Compare the sizes of the 8 groups of values ​​in the 4 groups of coordinates respectively, retain the smaller values ​​in xa1 and xb1, ya1 and yb1, xa2 and xb2, and ya4 and yb4, and record them as xc1, yc1, xc2, and yc4 respectively; retain the larger values ​​in ya2 and yb2, xa3 and xb3, ya3 and yb3, and xa4 and xb4, and record them as yc2, xc3, yc3, and xc4 respectively; Through the above steps, four new sets of coordinates are obtained, including: lower left corner coordinates (xc1, yc1), upper left corner coordinates (xc2, yc2), upper right corner coordinates (xc3, yc3), and lower right corner coordinates (xc4, yc4); The quadrilateral formed by the four sets of new coordinates is the new cotton boll marking frame of the overlapping image.

Images