A computer vision-based automatic sorting method and system for maize haploid

Abstract

The application discloses a kind of corn haploid automated sorting method and system based on computer vision;Among them, the method comprises: the image is preprocessed by cropping etc., three-face image is respectively input into corn haploid identification model, and classification result is obtained;If the classification results of three-face image are all non-embryo surface, it is considered that embryo part is not collected, and the screening air nozzle without collection information installed in the inside of seed metering disc is started, before the grain enters the seed pushing area, it is separated from suction hole, and returns to the seed filling area;If embryo surface is contained in three faces, this seed grain is separated from seed metering disc suction hole under the action of seed cleaning and seed scraping device, and is determined according to classification result;When haploid grain falls, seed sorting air nozzle starts and sorts it into haploid grain box, and diploid grain falls freely into diploid grain box when it is determined as diploid grain.The application does not need artificial placement of corn grain, makes the sorting process more automated, and is more suitable for image information collection of corn grain of various shapes such as round grain.

Description

Technical Field

[0001] This invention belongs to the field of maize breeding and non-destructive testing, specifically relating to an automated sorting method and system for maize haploids based on computer vision. Background Technology

[0002] Haploid breeding technology is a rapid and efficient breeding approach that can greatly shorten the breeding cycle. However, the natural induction rate of maize haploids is about 0.1%, and the artificial induction rate is only about 10%. Therefore, quickly and non-destructively separating haploid grains from a large number of grains is a key step in improving the efficiency of haploid breeding.

[0003] Currently, the most effective and reliable method for identifying haploids is the genetic marker method. This method involves introducing the Navajo marker gene R-nj into maize seeds. After being marked, haploid kernels only have a purple color mark in the endosperm, while diploid kernels have color marks in both the endosperm and the embryo. Therefore, haploid kernels can be quickly and non-destructively sorted by identifying whether there is a color mark in the embryo.

[0004] The automated sorting of haploid maize kernels uses two methods: machine vision and nuclear magnetic resonance (NMR). NMR identification is time-consuming and the equipment is expensive; machine vision is fast and the equipment is cheaper, but existing machine vision inspection equipment is cumbersome and bulky, requiring manual placement of maize kernels with the germ side facing up, making fully automated sorting impossible. It also has high requirements for the shape of the maize kernels, making it difficult to stably place and inspect spherical seeds, and is mostly only suitable for dent seeds.

[0005] Meanwhile, since the difference between haploid and diploid grains lies in the color markings on the embryo surface, the embryo must be placed face up manually when collecting images, so the efficiency of embryo photography needs to be improved urgently. Summary of the Invention

[0006] To address the problem of automated sorting of haploid corn kernels in the background art, this invention proposes a computer vision-based automated sorting method and system for corn haploids. This system automates the entire process of feeding, detection, and sorting. Furthermore, a single detection can obtain images of corn kernels from three different angles using mirrors mounted on both sides. This method is applicable not only to corn kernels of various shapes but also improves sorting efficiency. The technical solution of the method includes:

[0007] Step A: The haploid and diploid grains to be tested are fed into the air suction seed metering device through the feeding funnel. The grains are individually adsorbed onto the suction holes of the seed metering disc and rotate accordingly.

[0008] Step B: The kernels pass through the photoelectric sensor, triggering the color industrial camera to capture images. Using two plane mirrors installed in the image acquisition area, three images of the corn kernels passing through this area are captured simultaneously.

[0009] Step C: After preprocessing the image such as cropping, input the three images into the maize haploid recognition model to obtain the classification results;

[0010] Step D: If the classification results of the three images are all non-embryo surfaces, it is considered that the embryo has not been collected. The screening nozzle without collection information installed inside the seed metering tray is activated to remove the grain from the suction hole before it enters the seed pushing area and return to the seed filling area for the next round of image collection.

[0011] If the germ surface is included on three sides, the grain will be separated from the suction hole of the seed metering plate by the seed cleaning and scraping device and fall into the seed drop outlet. The classification results will be used to determine the grain. When the grain is determined to be haploid, the seed sorting nozzle will be activated to sort it into the haploid grain box. The grain is determined to be diploid and will fall freely into the diploid grain box.

[0012] The steps for establishing the maize haploid identification model are as follows:

[0013] Step 1: Put haploid and diploid corn kernels into the air suction seed metering device. The kernels are individually adsorbed onto the suction holes of the seed metering disc and pass through the image acquisition area in sequence. Two plane mirrors are installed on both sides here. A color industrial camera is vertically installed on the front shell of the seed metering device. One frame of image can simultaneously acquire image information of different sides of a corn kernel.

[0014] Step 2: Extract the region of interest, crop the original image, and obtain three single-sided images of a single seed from one frame;

[0015] Step 3: Classify and label the extracted single-sided images of the kernels into three categories: embryo haploid, embryo diploid, and non-embryonic. Perform data augmentation and size adjustment on them, and divide the images into training and test sets according to the proportions.

[0016] Step 4: Construct a convolutional neural network and input the training set into the convolutional neural network for training. Extract features and classify the training set to obtain a classification model after training.

[0017] Step 5: Use the test set as input to the convolutional neural network, evaluate the performance of the classification model based on the results, and thus obtain the maize haploid identification model.

[0018] The region of interest is selected based on the principle of plane mirror reflection imaging and the size of corn kernels, and the coordinates and size of the region of interest are determined.

[0019] The data augmentation method specifically includes flipping, rotating, and adjusting brightness to enhance the corn kernel dataset. The image size is adjusted using a double interpolation method, and the image dataset is divided into a training set and a test set in a 7:3 ratio.

[0020] The convolutional neural network includes convolutional layers, batch normalization layers, pooling layers, and fully connected layers, and uses a softmax classifier for classification.

[0021] At least 500 original images of haploid and diploid grains were collected, and 1,500 single-sided images of individual grains were cropped.

[0022] Each single-sided image of the grain was resized to 100*100 using bicubic interpolation.

[0023] Step 4, constructing the convolutional neural network, includes:

[0024] Step 41: Input image is 100*100*3;

[0025] Step 42: Four convolutional layers (CONV) using 5*5 kernels, with stride and padding both set to 1. The number of kernels is 32, 64, 128, and 128, respectively. The output structures of the convolutional layers are 98*98*32, 47*47*64, 21*21*128, and 8*8*128, respectively.

[0026] Each convolutional layer is followed by a batch normalization layer, and the output size remains unchanged. Batch normalization normalizes the output of each layer and then uses the parameters learned in BN to restore the data features before normalization. Batch normalization layers can speed up network training and convergence.

[0027] Four pooling layers (POOL) use max pooling operation, with a kernel size of 2*2;

[0028] Three fully connected layers (FC) with 2048, 512, and 128 features respectively are used. Dropout layers are used in the fully connected layers to avoid overfitting.

[0029] Step 43: Select ReLU as the activation function;

[0030] Step 44: The output layer uses a Softmax classifier to obtain three classifications;

[0031] Step 45: During training, the RMSprop optimizer and cross-entropy loss function are used. The initial learning rate is 0.0001. For every 100 iterations, the learning rate is reduced to 0.5 times the original value.

[0032] Step 46: Backpropagate the error, update the weights, and save the optimal model.

[0033] This invention also proposes an automated sorting system for maize haploids based on computer vision. The system's technical solution includes: a feeding hopper, a color industrial camera, haploid kernel boxes, diploid kernel boxes, a blower, kernel sorting nozzles, a pneumatic seed metering device, a seed cleaning and scraping component, and a non-information-collecting screening nozzle. The pneumatic seed metering device is vertically mounted above the frame. The feeding hopper is installed on one side of the pneumatic seed metering device and enters the seed filling area below it. Along the seed metering disc in the pneumatic seed metering device, the suction holes on the device are connected to the blower through the seed suction chamber to create a negative pressure environment. The kernel sorting nozzles and the non-information-collecting screening nozzles are connected to an air compressor through independent valves. A drive motor fixed outside the front housing is connected to the seed metering disc through a transmission structure inside the seed metering device. The haploid kernel boxes and diploid kernel boxes are mounted on the ground below the kernel sorting nozzles.

[0034] Along the rotation direction of the seed metering disc in the air suction seed meterer, there are sequentially arranged a seed filling area, a photoelectric sensor, an image acquisition area, a seed removal area without acquisition information, and a seed falling area with information. The photoelectric sensor is installed between the front housing and the seed metering disc, in front of the image acquisition area, and is used to externally trigger the camera.

[0035] Two plane mirrors fixed on a plane mirror bracket are installed in the image acquisition area. The color industrial camera is set on one side of the air suction seed metering device through the camera bracket. The lens of the color industrial camera focuses through the image acquisition port of the front housing and simultaneously faces the running trajectory of the two plane mirrors and the suction hole. The angle of the two plane mirrors allows the color industrial camera to simultaneously acquire images of the seeds on the suction hole from three directions.

[0036] The seed rejection area is equipped with a seed rejection nozzle. The seed rejection nozzle is aligned with the trajectory of the suction hole. If the embryo is not collected, the seed is sprayed back to the seed filling area for secondary collection.

[0037] The seed-collecting area is equipped with a seed-cleaning and scraping component. The upper end of the seed-cleaning and scraping component is close to the seed metering tray, scraping the seeds sucked on the suction holes away from the seed metering tray. The seeds are then discharged from the seed metering device through the seed-collecting port of the air-suction seed meterer. A seed sorting nozzle is installed outside the seed-collecting port. When the seeds are determined to be haploid, the seed sorting nozzle is activated to sort them into the haploid seed box. The seeds are determined to be diploid and fall freely into the diploid seed box.

[0038] Two plane mirrors are installed opposite each other on both sides of the suction hole, with their bottom edges parallel.

[0039] The size of the plane mirror is at least 16*20mm.

[0040] The non-information collection screening nozzle is set outside the air suction seed metering device. The side wall of the air suction seed metering device has a screening nozzle mounting hole. The nozzle of the non-information collection screening nozzle extends into the air suction seed metering device through the screening nozzle mounting hole.

[0041] The beneficial effects of this invention are as follows:

[0042] 1. Using an air-suction seed metering device to sort corn kernels into individual kernels and acquire images eliminates the need for manual placement of corn kernels, making the sorting process more automated and more suitable for acquiring image information of corn kernels of various shapes, such as round kernels.

[0043] 2. When collecting kernel images, this invention utilizes plane mirrors installed on both sides of the suction hole of the seed metering tray to collect three-sided image information of corn kernels at the same position using a camera. These three-sided images are then used as input to a classification model to comprehensively identify the properties of the corn kernels, thereby increasing the probability of collecting information about the embryo of the corn kernels and improving sorting efficiency.

[0044] 3. In the haploid sorting model, the collected maize images are divided into three categories: germ haploids, germ diploids, and non-germ haploids. If none of the three images contain germ information, they are considered non-germ haploids. The nozzles installed inside the seed metering disc are used to screen out the kernels for which germ information has not been collected, allowing them to fall back to the seed filling area of ​​the seed metering disc. They are then adsorbed onto the suction holes of the seed metering disc and images are collected again. For the same kernel, no secondary manual feeding is required.

[0045] 4. In convolutional neural network models, batch normalization layers are added after convolutional layers to speed up training and improve the network's generalization ability. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of an embodiment of an automated haploid sorting system for maize based on computer vision according to the present invention.

[0047] Figure 2 This is a schematic diagram of the internal structure of the seed metering device in an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the installation of the seed metering disc and the plane mirror in an embodiment of the present invention.

[0049] Figure 4 This is a flowchart of the sorting method in an embodiment of the present invention.

[0050] Figure 5 This is a schematic diagram of the model structure of the sorting method in an embodiment of the present invention.

[0051] Among them: 1-feeding funnel, 2-color industrial camera, 3-camera bracket, 4-air compressor, 5-haploid grain box, 6-diploid grain box, 7-fan, 8-grain sorting nozzle, 9-air suction seed metering device, 10-screening nozzle without data collection, 901-front housing, 902-screening nozzle mounting hole, 903-photoelectric sensor, 904-plane mirror, 905-seed suction chamber, 906-drive motor, 907-seed cleaning and scraping component, 908-seed metering disc, 909-seed drop port, 910-suction hole, 911-protrusion. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 The embodiment of the present invention shown includes a system comprising: a feeding hopper 1, a color industrial camera 2, a haploid grain box 5, a diploid grain box 6, a blower 7, a grain sorting nozzle 8, a pneumatic seed metering device 9, a seed cleaning and scraping component 907, and a screening nozzle 10 for non-collection information. The pneumatic seed metering device 9 is vertically mounted above the frame. The feeding hopper 1 is installed on one side of the pneumatic seed metering device 9 and enters the seed filling area below the pneumatic seed metering device 9. The seed metering disc 908 and the suction holes 910 on the air suction seed meterer 9 are connected to the fan 7 through the seed suction air chamber 905 to create a negative pressure environment; the grain sorting nozzle 8 and the non-collection information screening nozzle 10 are connected to the air compressor 4 through independent valves; the drive motor 906, which is fixed outside the front housing 901, is connected to the seed metering disc 908 through the internal transmission structure of the seed meterer; the haploid grain box 5 and the diploid grain box 6 are installed on the ground below the grain sorting nozzle 8.

[0054] Along the rotation direction of the seed metering disc 908 in the air suction seed meterer 9, there are sequentially arranged a seed filling area, a photoelectric sensor, an image acquisition area, a seed removal area without acquisition information, and a seed falling area with information. The photoelectric sensor is installed between the front housing and the seed metering disc, in front of the image acquisition area, and is used to externally trigger the camera.

[0055] Two plane mirrors 904 fixed on a plane mirror bracket are installed in the image acquisition area. The color industrial camera 2 is set on one side of the air suction seed metering device 9 through the camera bracket 3. The lens of the color industrial camera 2 focuses through the image acquisition port of the front housing 901 and simultaneously faces the running trajectory of the two plane mirrors 904 and the suction hole 910. The angle of the two plane mirrors 904 allows the color industrial camera 2 to simultaneously acquire images of the seeds on the suction hole 910 from three directions.

[0056] The seed rejection area is equipped with a seed rejection nozzle 10. The seed rejection nozzle 10 is aligned with the running trajectory of the suction hole 910. When the embryo is not collected, the seed is sprayed back to the seed filling area for secondary collection.

[0057] A seed cleaning and scraping component 907 is provided at the seed-dropping area. The upper end of the seed cleaning and scraping component 907 is close to the seed metering tray 908, scraping the seeds sucked on the suction hole 910 away from the seed metering tray 908, and then discharged from the seed-dropping port 909 of the air-suction seed meterer 9. A seed sorting nozzle 8 is installed outside the seed-dropping port 909. When the seeds are determined to be haploid, the seed sorting nozzle is activated to sort them into the haploid seed box 5. The seeds are determined to be diploid and fall freely into the diploid seed box 6 (the opening time and size of the seed sorting nozzle 8 depend on the installation position of the haploid seed box 5 and the diploid seed box 6).

[0058] In this embodiment, based on the size of the corn kernels, in order to acquire a more complete image of the kernels, two plane mirrors are installed opposite each other on both sides of the suction hole, with parallel bottom edges and an inclined installation. The size of the plane mirrors is at least 16*20mm. The purpose is to enable the camera to not only acquire the actual image of the kernel facing the lens, but also to acquire the mirror image in the two plane mirrors, increasing the possibility of acquiring germ surface information. The plane mirrors are fixed in the plane mirror bracket, which is fixed inside the air suction seed metering device 9. In order to make the side image of the corn kernels acquired by the plane mirrors more complete, the seed metering plate is a seed metering plate with protrusions 911 around the suction hole 910, and a groove is provided at the mirror mounting position.

[0059] In this embodiment, in order to prevent the start and stop of the air nozzle from interfering with the seed suction effect inside the air suction seed metering device 9, the non-collection information screening air nozzle 10 is set outside the air suction seed metering device 9. The side wall of the air suction seed metering device 9 has a screening air nozzle mounting hole 902, and the nozzle of the non-collection information screening air nozzle 10 extends into the air suction seed metering device 9 from the screening air nozzle mounting hole 902.

[0060] The system's sorting method is as follows:

[0061] 1) The haploid and diploid grains to be tested are fed into the air suction seed metering device 9 through the feed funnel 1. The drive motor 906 drives the seed metering disc 908 to rotate. Under the disturbance of the seed metering disc and the action of the negative pressure air chamber 905, the grains are individually adsorbed onto the suction holes of the seed metering disc and rotate accordingly.

[0062] 2) When the kernel passes through the photoelectric sensor 903, it triggers the color industrial camera 2 to capture a picture. Using the two plane mirrors 904 installed in the image acquisition area, three-sided images of the corn kernel passing through this area are captured at the same time.

[0063] 3) After preprocessing the images, such as cropping, the three images are input into the maize haploid recognition model to obtain the classification results;

[0064] 4) If the classification results of the three images are all non-embryonic surfaces, it is considered that the embryo has not been collected. The screening nozzle without collection information installed inside the seed metering tray is activated to remove the grain from the suction hole before it enters the seed pushing area and return to the seed filling area for the next round of image collection.

[0065] If the germ surface is included on three sides, the grain will be separated from the suction hole of the seed metering plate under the action of the seed cleaning and scraping device 907 and fall into the seed drop outlet 909. The classification result will be used to determine the grain. When the grain is determined to be haploid, the seed sorting nozzle will be activated to sort it into the haploid grain box 5. The grain is determined to be diploid and falls freely into the diploid grain box 6.

[0066] The system specifically uses a sorting method based on a constructed maize haploid identification model. The steps for establishing the maize haploid identification model are as follows:

[0067] Step 1: Put haploid and diploid corn kernels into the air suction seed metering device 9. The kernels are individually adsorbed onto the suction holes of the seed metering disc 908 and pass through the image acquisition area in sequence. Two plane mirrors 904 are installed on both sides here. The color industrial camera 2 is vertically installed on the front housing 901 of the seed metering device. One frame of image can simultaneously acquire image information of three different sides of a corn kernel.

[0068] Step 2: Extract the region of interest (ROI). Crop the original image to obtain three single-sided images of individual kernels from one frame. The ROI selection is based on the principle of plane mirror reflection imaging and the size of the corn kernels. Determine the coordinates and size of the ROI.

[0069] Step 3: Classify and label the extracted single-sided images of the kernels into three categories: haploid embryo, diploid embryo, and non-embryo. Perform data augmentation and size adjustment on the kernels, and divide the images into training and test sets according to the ratio. The data augmentation methods include flipping, rotating, and adjusting brightness to enhance the corn kernel dataset. The image size is adjusted by double interpolation, and the image dataset is divided into training and test sets in a 7:3 ratio.

[0070] Step 4: Construct a convolutional neural network and input the training set into the convolutional neural network for training. Perform feature extraction and classification on the training set to obtain a classification model after training. The convolutional neural network includes convolutional layers, batch normalization layers, pooling layers, and fully connected layers, and uses a softmax classifier for classification.

[0071] Step 5: Use the test set as input to the convolutional neural network, evaluate the performance of the classification model based on the results, and thus obtain the maize haploid identification model;

[0072] In order to collect multi-faceted images of corn kernels, the sample kernels can be fed into the seed metering device twice or more for image acquisition. At least 500 original images of haploid and diploid kernels should be collected, and 1,500 single-faced images of individual kernels can be obtained by cropping.

[0073] Each single-sided image of the grain was resized to 100*100 using bicubic interpolation.

[0074] Step 4, constructing the convolutional neural network, includes:

[0075] Step 41: Input image is 100*100*3;

[0076] Step 42: Four convolutional layers (CONV) using 5*5 kernels, with stride and padding both set to 1. The number of kernels is 32, 64, 128, and 128, respectively. The output structures of the convolutional layers are 98*98*32, 47*47*64, 21*21*128, and 8*8*128, respectively.

[0077] Each convolutional layer is followed by a batch normalization layer (BN). The output size remains unchanged. Batch normalization normalizes the output of each layer and then uses the parameters learned in BN to restore the data features before normalization. Batch normalization layers can speed up network training and convergence.

[0078] Four pooling layers (POOL) use max pooling operation, with a kernel size of 2*2;

[0079] Three fully connected layers (FC) with 2048, 512, and 128 features respectively are used. Dropout layers are used in the fully connected layers to avoid overfitting.

[0080] Step 43: Select ReLU as the activation function;

[0081] Step 44: The output layer uses a Softmax classifier to obtain three classifications;

[0082] Step 45: During training, the RMSprop optimizer and cross-entropy loss function are used. The initial learning rate is 0.0001. For every 100 iterations, the learning rate is reduced to 0.5 times the original value.

[0083] Step 46: Backpropagate the error, update the weights, and save the optimal model.

Claims

1. A computer vision-based automated sorting method for maize haploids, characterized in that, include: Step A: The haploid and diploid grains to be tested are fed into the air suction seed metering device through the feeding funnel. The grains are individually adsorbed onto the suction holes of the seed metering disc and rotate accordingly. Step B: The kernels pass through the photoelectric sensor, triggering the color industrial camera to capture images. Using two plane mirrors installed in the image acquisition area, three images of the corn kernels passing through this area are captured simultaneously. Step C: After preprocessing the image by cropping, input the three images into the maize haploid recognition model to obtain the classification results; Step D: If the classification results of the three images are all non-embryo surfaces, it is considered that the embryo has not been collected. The screening nozzle without collection information installed inside the seed metering tray is activated to remove the grain from the suction hole before it enters the seed pushing area and return to the seed filling area for the next round of image collection. For the same grain, no secondary manual feeding is required. If the germ surface is included on three sides, the grain will be separated from the suction hole of the seed metering plate by the seed cleaning and scraping device and fall into the seed drop outlet. The classification results will be used to determine the grain. When the grain is determined to be haploid, the seed sorting nozzle will be activated to sort it into the haploid grain box. The grain is determined to be diploid and will fall freely into the diploid grain box. The steps for establishing the maize haploid identification model are as follows: Step 1: Put haploid and diploid corn kernels into the air suction seed metering device. The kernels are individually adsorbed onto the suction holes of the seed metering disc and pass through the image acquisition area in sequence. Two plane mirrors are installed on both sides here. A color industrial camera is vertically installed on the front shell of the seed metering device. One frame of image simultaneously acquires image information of three different sides of a corn kernel. Step 2: Extract the region of interest, crop the original image, and obtain three single-sided images of a single seed from one frame; Step 3: Classify and label the extracted single-sided images of the kernels into three categories: embryo haploid, embryo diploid, and non-embryonic. Perform data augmentation and size adjustment on them, and divide the images into training and test sets according to the proportions. Step 4: Construct a convolutional neural network and input the training set into the convolutional neural network for training. Extract features and classify the training set to obtain a classification model after training. Step 5: Use the test set as input to the convolutional neural network, evaluate the performance of the classification model based on the results, and thus obtain the maize haploid identification model.

2. The automated sorting method for maize haploids based on computer vision according to claim 1, characterized in that, The region of interest is selected based on the principle of plane mirror reflection imaging and the size of corn kernels, and the coordinates and size of the region of interest are determined.

3. The automated sorting method for maize haploids based on computer vision according to claim 1, characterized in that, The data augmentation method specifically includes flipping, rotating, and adjusting brightness to enhance the corn kernel dataset. The image size is adjusted using a double interpolation method, and the image dataset is divided into a training set and a test set in a 7:3 ratio.

4. The automated sorting method for maize haploids based on computer vision according to claim 1, characterized in that, The convolutional neural network includes convolutional layers, batch normalization layers, pooling layers, and fully connected layers, and uses a softmax classifier for classification.

5. The automated sorting method for maize haploids based on computer vision according to claim 1, characterized in that, At least 500 original images of haploid and diploid grains were collected, and 1,500 single-sided images of individual grains were cropped.

6. The automated sorting method for maize haploids based on computer vision according to claim 1, characterized in that, Each single-sided image of the grain was resized to 100*100 using bicubic interpolation. Step 4, constructing the convolutional neural network, includes: Step 41: Input image is 100*100*3; Step 42: Four convolutional layers using 5*5 kernels, with stride and padding of 1. The number of kernels is 32, 64, 128, and 128 respectively. The output structures of the convolutional layers are 98*98*32, 47*47*64, 21*21*128, and 8*8*128 respectively. Each convolutional layer is followed by a batch normalization layer, and the output size remains unchanged. Batch normalization normalizes the output of each layer and then uses the parameters learned in BN to restore the data features before normalization. Batch normalization layers can speed up network training and convergence. Four pooling layers, using max pooling operation, with a kernel size of 2*2; Three fully connected layers with 2048, 512, and 128 features respectively are used. Dropout layers are used in the fully connected layers to avoid overfitting. Step 43: Select ReLU as the activation function; Step 44: The output layer uses a Softmax classifier to obtain three classifications; Step 45: During training, the RMSprop optimizer and cross-entropy loss function are used. The initial learning rate is 0.0001. For every 100 iterations, the learning rate is reduced to 0.5 times the original value. Step 46: Backpropagate the error, update the weights, and save the optimal model.

7. An automated sorting system based on the computer vision-based automated sorting method for haploid maize as described in claim 1, characterized in that, include: The machine includes a feed hopper (1), a color industrial camera (2), a haploid grain box (5), a diploid grain box (6), a blower (7), a grain sorting nozzle (8), a pneumatic seed metering device (9), a seed cleaning and scraping component (907), and a screening nozzle without data collection (10). The pneumatic seed metering device (9) is vertically mounted above the frame. The feed hopper (1) is installed on one side of the pneumatic seed metering device (9) and enters the seed filling area below the pneumatic seed metering device (9). The feed hopper (1) is positioned along the seed metering disc (908) in the pneumatic seed metering device (9). The suction hole (910) on the air suction seed meter (9) is connected to the blower (7) through the seed suction air chamber (905) to create a negative pressure environment; the seed sorting nozzle (8) and the non-collection information screening nozzle (10) are connected to the air compressor (4) through independent valves; the drive motor (906) fixed outside the front housing (901) is connected to the seed metering disc (908) through the transmission structure inside the seed meter; the haploid seed box (5) and the diploid seed box (6) are installed on the ground below the seed sorting nozzle (8); Along the rotation direction of the seed metering disc (908) in the air suction seed meterer (9), there are a seed filling area, a photoelectric sensor, an image acquisition area, a seed removal area without acquisition information, and a seed falling area with information. The photoelectric sensor is installed between the front housing and the seed metering disc, in front of the image acquisition area, and is used to trigger the camera externally. Two plane mirrors (904) fixed on a plane mirror bracket are installed at the position of the image acquisition area. The color industrial camera (2) is set on one side of the air suction seed meter (9) through the camera bracket (3). The lens of the color industrial camera (2) focuses through the image acquisition port of the front housing (901) and simultaneously faces the running trajectory of the two plane mirrors (904) and the suction hole (910). The angle of the two plane mirrors (904) allows the color industrial camera (2) to simultaneously acquire images of the seeds on the suction hole (910) in three directions. A no-collection-information-information-screening nozzle (10) is provided at the location of the no-collection-information-information-removal area. The no-collection-information-information-screening nozzle (10) is directly aligned with the running trajectory of the suction hole (910). When the embryo is not collected, the seed is sprayed back to the filling area for secondary collection. A seed cleaning and scraping component (907) is provided at the seed dropping area. The upper end of the seed cleaning and scraping component (907) is close to the seed metering tray (908) to scrape the seeds sucked on the suction hole (910) away from the seed metering tray (908) and then discharged from the seed dropping port (909) of the air suction seed meterer (9). A seed sorting nozzle (8) is installed outside the seed dropping port (909). When the seeds are determined to be haploid, the seed sorting nozzle is activated to sort them into the haploid seed box (5). The seeds are determined to be diploid and fall freely into the diploid seed box (6).

8. The automated haploid sorting system for maize based on computer vision according to claim 7, characterized in that, Two plane mirrors are installed opposite each other on both sides of the suction hole, with their bottom edges parallel.

9. The automated haploid sorting system for maize based on computer vision according to claim 8, characterized in that, The size of the plane mirror is at least 16*20mm.

10. The automated haploid sorting system for maize based on computer vision according to claim 7, characterized in that, The non-information collection screening nozzle (10) is set outside the air suction seed meter (9). The air suction seed meter (9) has a screening nozzle mounting hole (902) on its side wall. The nozzle of the non-information collection screening nozzle (10) extends into the air suction seed meter (9) from the screening nozzle mounting hole (902).

Images