A Wheat Spike Number Recognition Method, System, and Medium Based on Faster R-CNN

Abstract

The present invention belongs to the field of intelligent biotechnology and relates to a method, system and medium for wheat spike number recognition based on Faster R-CNN, including: extracting the canopy images of wheat spikes of different families, calibrating the wheat spikes in the images to obtain the labels corresponding to each wheat spike; inputting the images and labels of the wheat spikes into a ResNet network model for feature extraction to obtain the feature map of the wheat spikes, and establishing corresponding candidate boxes according to the wheat spike contours; training the Faster R-CNN model through the feature map to obtain an optimal wheat spike recognition model; inputting the wheat spike image to be measured into the wheat spike recognition model to obtain the candidate boxes corresponding to each wheat spike; counting the number of candidate boxes, so as to obtain the number of wheat spikes. It can accurately perform QTL mapping on wheat, can accurately identify wheat with high overlapping characteristics of wheat spike number, and is expected to provide a high-throughput analysis tool for the identification of yield-related phenotypes in wheat molecular breeding.

Description

Technical Field

[0001] The present invention relates to a method, system and medium for identifying the number of wheat ears based on Faster R-CNN, belonging to the field of intelligent biotechnology, and particularly to the field of automatic wheat ear number identification technology. Background Art

[0002] As an important staple food crop in China, the morphological parameters of the wheat ear directly reflect the growth status and yield information of wheat (Yang Jinwen et al., 2013), and are important parameters reflecting the quality and yield of wheat. At present, the statistics of the number of spikelets and grains in the wheat ear mainly rely on manual counting, which is not only time-consuming and laborious, but also inefficient. The existing means to solve this problem mainly include analyzing plant phenotypes through optical imaging methods, which have obvious advantages compared with manual counting. However, for wheat, due to the overlapping of adjacent leaves, or the overlapping of ears and fruits, resulting in a large number of occlusions, it is difficult for optical imaging to obtain accurate wheat ear numbers.

[0003] At present, deep learning models, with a framework similar to the interconnection and operation between human neurons, summarize the high-level abstract laws in data features through feature learning of a large amount of data, and have made great progress in many different fields. In particular, deep learning methods such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have achieved great success in the fields of image classification, object recognition and sequence feature extraction. Therefore, at this stage, some scholars have also applied neural network models to crop feature recognition, such as using image processing technology and deep learning methods to segment crops from the background, so as to obtain a series of important phenotypic parameters such as the length and width of wheat ears, the projected area and orientation of grains wrapped in glumes; in the study of maize ears, important ear phenotypic parameter information such as the number of rows, grains per row, and total number of grains has been extracted through image processing; in the study of rice ears, high-precision extraction of rice ear structure features and phenotypic information related to the number of grains has been achieved.

[0004] However, due to the complex structure of the wheat ear, the small plant spacing, the high overlap of wheat ears, and the presence of awns in some wheat varieties, it is difficult to accurately identify the phenotypes of wheat ears. The existing wheat ear identification methods can only analyze the length and width of wheat ears, the projected area and orientation of grains wrapped in glumes, etc., that is, they can only analyze the relatively clear and non-overlapping parts of the wheat ear, and cannot be applied to the overall study of wheat ears, such as the number of wheat ears. Summary of the Invention

[0005] In view of the above problems, the object of the present invention is to provide a wheat spike number recognition method, system and medium based on Faster R-CNN, which can accurately perform QTL mapping on wheat, can accurately identify wheat with wheat spike numbers having high overlap characteristics, and is expected to provide a high-throughput analysis tool for the identification of yield-related phenotypes in wheat molecular breeding.

[0006] To achieve the above object, the present invention proposes the following technical solutions: A wheat spike number recognition method based on Faster R-CNN includes: extracting the canopy images of wheat spikes of different families, calibrating the wheat spikes in the images to obtain the labels corresponding to each wheat spike; inputting the images and labels of the wheat spikes into the ResNet network model for feature extraction to obtain the feature map of the wheat spikes, and establishing corresponding candidate boxes according to the wheat spike contours; training the Faster R-CNN model through the feature map to obtain the optimal wheat spike recognition model; inputting the image of the wheat spike to be measured into the wheat spike recognition model to obtain the candidate boxes corresponding to each wheat spike; counting the number of candidate boxes to obtain the number of wheat spikes.

[0007] Furthermore, QTL mapping is performed on the genetic traits of wheat through the number of wheat spikes per unit area.

[0008] Furthermore, the QTL mapping method includes: obtaining a double haploid population of wheat; genotyping the double haploid population of wheat through an SNP chip; inputting the genotyping results into the IciMapping network, deleting redundant markers, and constructing a genetic linkage map; obtaining the QTL mapping of the wheat genetic traits through composite interval mapping based on the genetic linkage map.

[0009] Furthermore, the genotyping method includes: performing whole-genome amplification on the genomic DNA of the wheat to be measured to obtain an amplification product; cutting the amplification product with a random endonuclease to obtain DNA fragments; hybridizing the DNA fragments with an SNP chip so that the DNA fragments are complementary to the specific capture probes on the beads of the SNP chip; washing to remove the unhybridized or mismatched hybridized DNA fragments; performing single-base extension on the specific capture probes by using dinitrophenol and biotin-labeled nucleotide substrates, and staining so that different nucleotide substrates are labeled with different fluorescent dyes; scanning the SNP chip, and judging and outputting the genotyping results according to the detected fluorescence counts.

[0010] Furthermore, training the Faster R-CNN model includes adjusting the ratio of the candidate boxes, the IOU threshold, and optimizing the strategy learning rate of model training.

[0011] Further, through L2 regularization of the fully connected layer and softmax classification and regression, it is determined whether the candidate bounding box exceeds the threshold to determine whether the candidate bounding box contains a complete wheat ear. For the candidate bounding boxes that exceed the threshold, their position coordinates are obtained through softmax regression, and the candidate bounding boxes that are greater than the IOU threshold but are not the optimal prediction for the true wheat ear are removed through non-maximum suppression.

[0012] Further, the images and labels of the wheat ears are input into several sub-modules of the ResNet network model for feature extraction. Candidate bounding boxes are generated based on the extracted features. The generated candidate bounding boxes, together with the original images and labels of the wheat ears, are subjected to L2 regularization and softmax classification and regression. The intersection over union (IoU) of their results is calculated, and non-maximum suppression is performed on the results to obtain a preliminary classification scheme. The preliminary classification scheme is successively input into the ROI pooling layer, another sub-module of the ResNet network model, L2 regularization, the softmax layer, IoU calculation, and non-maximum suppression, thereby outputting the final image recognition result.

[0013] Further, several sub-modules of the ResNet network model include four sub-modules. The first sub-module includes a convolutional layer, a BN layer, a RELU layer, and a max pooling layer. The second to fourth sub-modules each include a convolutional block and several recognition blocks. Another sub-module of the ResNet network model includes a convolutional block, two recognition blocks, a max pooling layer, and a flattening layer.

[0014] The present invention also discloses a wheat ear number recognition system based on Faster R-CNN, including: a calibration module for extracting the wheat ear canopy images of different families and calibrating the wheat ears in the images to obtain the labels corresponding to each wheat ear; a feature map establishment module for inputting the images and labels of the wheat ears into the ResNet network model for feature extraction to obtain the feature map of the wheat ears and establishing corresponding candidate bounding boxes according to the wheat ear contours; a model training module for training the model based on the extracted feature map to obtain the optimal wheat ear recognition model; a wheat ear recognition module for inputting the wheat ear image to be measured into the wheat ear recognition model to obtain the candidate bounding boxes corresponding to each wheat ear; and a wheat ear number output module for counting the number of candidate bounding boxes to obtain the number of wheat ears.

[0015] The present invention also discloses a computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to execute the wheat ear number recognition method based on Faster R-CNN according to any one of the above.

[0016] Due to the above technical solutions adopted by the present invention, it has the following advantages:

[0017] 1. The present invention establishes a phenotypic recognition model for the number of spikes per unit area of wheat, and locates QTL loci related to yield traits based on this model. Experiments show that the model established by the present invention can quickly and efficiently identify the number of spikes per unit area of wheat families to be tested. Combining with the genotype data of the wheat to be tested can assist in screening high-yield families, laying a theoretical foundation for breeding wheat varieties with high yield, stable yield and excellent quality.

[0018] 2. The present invention organically combines molecular marker-assisted selection breeding with deep learning object detection, and establishes a method that can be used to assist breeders in selecting high-yield families. Using the model of the present invention, phenotypic data of the number of spikes per unit area of different families can be obtained in a high-throughput manner, and the high-yield potential of the wheat population to be tested can be quickly and accurately identified according to the genotype, accelerating the genetic gain of wheat yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic diagram of the wheat spike number recognition method based on Faster R-CNN in an embodiment of the present invention;

[0020] Figure 2 is a visualization image of the loss function in the training process of the wheat ear recognition model in an embodiment of the present invention;

[0021] Figure 3 is a high-quality RGB image of the wheat ear canopy of different families or varieties in an embodiment of the present invention, Figure 3 (a) is the original wheat ear canopy image, Figure 3 (b) is the cropped wheat ear canopy image, Figure 3 (c) is the labeled wheat ear canopy image;

[0022] Figure 4 is a schematic diagram of the wheat ear number output result in an embodiment of the present invention, Figure 4 (a) is the original wheat ear canopy image, Figure 4 (b) is the finally output wheat ear canopy image;

[0023] Figure 5 is a QTL mapping analysis chart of the number of spikes per unit area obtained by three methods of MSN, ISN and VSN in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0024] In order to enable those skilled in the art to better understand the technical direction of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the provision of the specific embodiments is only for better understanding of the present invention, and they should not be construed as limitations to the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description, and cannot be construed as indicating or implying relative importance.

[0025] Zhongmai 895 is a semi-winter multi-ear type mid-late maturing variety bred by cross-breeding Zhoumai 16 as the female parent and Likeng 4 as the male parent by the Institute of Crop Sciences, Chinese Academy of Agricultural Sciences and the Cotton Research Institute, Chinese Academy of Agricultural Sciences. It showed characteristics such as high yield and wide adaptability, and high temperature tolerance in the late filling stage in the variety comparison test and field demonstration from 2013 to 2015. Yangmai 16 is the variety with the largest planting area in the middle and lower reaches of the Yangtze River wheat region, and has characteristics such as fast filling speed and high grain weight.

[0026] The present invention relates to a method, system and medium for identifying the number of wheat ears based on Faster R-CNN. Taking 101 families of a doubled haploid (DH) population created with Yangmai 16 and Zhongmai 895 as parents as the object, the phenotypic identification and detection of the number of wheat ears are carried out based on the Faster-RCNN model in deep learning. This wheat ear identification method can accurately verify the number of wheat ears, and combine 660K SNP chip data for molecular marker-assisted breeding of the number of ears per unit area, and perform QTL mapping on the phenotypic data of the number of ears per unit area of wheat. The results show that this model can be used to identify or assist in identifying the number of ears per unit area of wheat, and thus is expected to provide a high-throughput analysis tool for the phenotypic identification of wheat molecular breeding and yield-related phenotypes. The following combines the accompanying drawings to detail the solution of the present invention through three embodiments.

[0027] Embodiment 1

[0028] A method for identifying the number of wheat ears based on Faster R-CNN in this embodiment is as Figure 1 shown, including:

[0029] S1 Extract the ear canopy images of different families, and calibrate the wheat ears in the images to obtain the labels corresponding to each wheat ear.

[0030] Please refer to Figure 3 as Figure 3 (a) shows, this embodiment provides 1840 high-quality ear canopy RGB images of wheat. Among them, 1032 images come from 166 natural populations and are used as the training set of the model to train the model. 808 images come from the DH population created with Yangmai 16 and Zhongmai 895 as parents and are used as the validation set of the model to train the model and perform subsequent QTL mapping. As Figure 3As shown in (b), a square frame of 0.5×0.5 cm was randomly placed in the non-marginal area of the plot as the object for taking pictures, so as to ensure the consistency of the object area obtained each time and perform cropping operations on the pictures. The above RGB images of the wheat ear canopy were taken in the middle stage of wheat filling. While extracting the wheat ear canopy images of different families, the number of wheat ears in the square frame of 0.5×0.5 cm was counted manually at the same time. Each family or variety was counted twice and the average value was taken for later verification. The phenotypic results of this part are represented by MSN (manual field phenotypic statistical data).

[0031] As Figure 3 As shown in (c), through the Labelimg software, the heads of the wheat ears in each obtained wheat ear canopy image were calibrated, and the number of labels appearing in each image was counted. The phenotypic results of this part are represented by ISN (data annotated based on the Labelimg software). During the calibration process, overlapping and occluded wheat ears should also be calibrated to ensure the accuracy of model training.

[0032] S2 Input the images and labels of the wheat ears into the ResNet50 backbone feature extraction network model for feature extraction to obtain the feature maps of the wheat ears, and establish corresponding candidate boxes according to the wheat ear contours.

[0033] In this embodiment, the original features were extracted with a stride of 8. After continuous downsampling convolution by the ResNet50 network model, feature maps with higher dimensions and more texture features were obtained.

[0034] The RPN network generates candidate boxes of different sizes and ratios in the feature maps, where the size is the area ratio of the candidate box. Each candidate box includes 4 area ratios and has three square boxes of different area sizes. Considering the characteristics of dense and small wheat ears, in this embodiment, one pixel point of each feature map is set to include 12 generated candidate boxes. The aspect ratio of each candidate box can be selected as 0.5, 1.0 or 2.0, and the size is 0.25, 0.5, 1.0 or 2.0, so that each wheat ear is included in the candidate box. The IOU thresholds are set to 0.3 and 0.7, which are used as the basis for discarding and adjusting the coordinates of the generated candidate boxes. When the IOU is less than 0.3, the candidate box is discarded. When the IOU is greater than 0.7, the generated candidate box is retained, and the position of the candidate box is adjusted through softmax regression to make it closer to the actual situation of the wheat ear.

[0035] S3 Train the Faster R-CNN model through the feature maps to obtain the optimal wheat ear recognition model;

[0036] As Figure 1As shown in the figure, Faster R-CNN mainly includes four parts: the feature extraction module (fully connected layer), RPN (Region Proposal Networks), ROI pooling layer, and the last fully connected layer that can be used for classification and regression. Compared with other two-stage models, Faster R-CNN uses the RPN structure, which greatly reduces the time used to extract candidate boxes and makes it easier to connect the extracted candidate boxes with the subsequent network structure as a whole, so as to be able to more accurately and quickly locate and classify the objects to be detected. In this embodiment, the feature extraction module of the ResNet50 residual model and the weights trained on the VOC2007 dataset are used as the initial weights for wheat ear recognition, so that the model can obtain the weight combination with the minimum loss function faster.

[0037] Training the Faster R-CNN model includes adjusting the ratio of candidate boxes, the IOU threshold, and optimizing the learning rate of the model training strategy.

[0038] Through the L2 regularization of the fully connected layer and softmax classification and regression, it is judged whether the candidate box exceeds the threshold to determine whether the candidate box contains a complete wheat ear. For the candidate boxes that exceed the threshold, their position coordinates are obtained through softmax regression. Here, 0.3 is used as the threshold to judge whether the candidate box contains a wheat ear, and the candidate boxes with an IOU threshold greater than 0.7, but not the optimal prediction for the real wheat ear, are removed through non-maximum suppression. According to the number of wheat ears in the wheat ear canopy image being approximately between 100 and 250, the number of wheat ears finally retained after non-maximum suppression is set to 300.

[0039] Using the Faster R-CNN model, combined with the morphological and density characteristics of wheat ears in the field, some parameters in Faster R-CNN are adjusted to adapt to the recognition of wheat ears, and the model is trained with the pictures obtained from the natural population as the training set. The specific parameter changes include adjusting the ratio of candidate boxes, adjusting the threshold of IOU (intersection-over-unio), and the learning rate of the model training strategy in three parts:

[0040] Taking the learning rate of 3e -4 as the initial value, after the number of iterations reaches 90,000 times, the learning rate is reduced by one-tenth, and the wheat ear recognition model is trained with a momentum of 0.9 and 500,000 iterations.

[0041] S4 inputs the wheat ear image to be measured into the wheat ear recognition model to obtain candidate boxes corresponding to each wheat ear;

[0042] S5 counts the number of candidate boxes to obtain the number of wheat ears, and its output result is as Figure 4 shown.

[0043] The specific structure of the Faster R-CNN model is as follows Figure 3 shown. The image and label of the wheat ear are input into several sub-modules of the ResNet network model for feature extraction. Candidate boxes are generated based on the extracted features. The generated candidate boxes, together with the original wheat ear image and label, are subjected to L2 regularization and softmax classification and regression. The intersection over union (IoU) of the results is calculated, and non-maximum suppression is performed on the results to obtain a preliminary classification scheme. The preliminary classification scheme is successively input into the ROI pooling layer, another sub-module of the ResNet50 network model, L2 regularization, the softmax layer, IoU calculation, and non-maximum suppression, so as to output the final image recognition result.

[0044] Several sub-modules of the ResNet network model include four sub-modules. The first sub-module includes a convolutional layer, a BN layer, a RELU layer, and a max pooling layer; the second module includes convolutional blocks and two recognition blocks; the third module includes convolutional blocks and three recognition blocks; the fourth module includes convolutional blocks and four recognition blocks; another sub-module of the ResNet50 network model includes convolutional blocks, two recognition blocks, a max pooling layer, and a flattening layer.

[0045] For the wheat ear recognition model, the model performance is mainly evaluated from two aspects: the loss function and the generalization ability. The visualization image of the loss function of the wheat ear recognition model in this embodiment is as follows Figure 2 shown; the generalization ability uses accuracy, precision, recall, and F1 score as the evaluation indicators of the wheat ear recognition model, and the results are shown in Table 1.

[0046] Table 1 Evaluation index table of the generalization ability of the wheat ear recognition model

[0047]

[0048]

[0049] The above model can quickly and accurately evaluate the number of ears per unit area. In this embodiment, 50 wheat ear canopy images are randomly verified. Compared with the artificial phenotype data MSN, the average accuracy is 86.7%. For comparison, in this embodiment, the average accuracy is 50% and 83% respectively according to the model output method (VSN) and the image annotation method (ISN) compared with the artificial phenotype data MSN. It can be seen that the method in this application has higher accuracy compared with the model output method (VSN) and the image annotation method (ISN).

[0050] Using SAS 9.4 software, basic statistics, calculation of broad-sense heritability, and variance analysis were performed on the phenotypes obtained by the MSN, ISN, and VSN methods. The results showed that the broad-sense heritability of the spike number per unit area phenotypes obtained by the above three methods was relatively high, all between 0.71 and 0.93, the phenotypic coefficient of variation was between 11.2% and 13.4%, and all conformed to the normal distribution, being suitable for QTL mapping. QTL mapping of the genetic traits of wheat was carried out through the spike number in unit area. The methods for QTL mapping include:

[0051] Obtaining a doubled haploid population of wheat.

[0052] Using the Illumina SNP genotyping platform to perform 660K SNP chip genotyping on the DH population of Yangmai 16 / Zhongmai 895, including BS, BobWhite, CAP, D_contig, etc., a total of 630,518. The doubled haploid population of wheat was genotyped by the SNP chip. The genotyping methods include: performing whole-genome amplification on the genomic DNA of the wheat to be tested to obtain an amplification product; cutting the amplification product with a random endonuclease to obtain DNA fragments; hybridizing the DNA fragments with the SNP chip to make the DNA fragments complementarily bind to the specific capture probes on the beads of the SNP chip; washing to remove the unhybridized or mismatched hybridized DNA fragments; performing single-base extension on the specific capture probes through dinitrophenol and biotin-labeled nucleotide substrates, and through staining, making different nucleotide substrates label different fluorescent dyes; scanning the SNP chip, and judging and outputting the genotyping results according to the detected fluorescence counts.

[0053] By inputting the genotyping results into the IciMapping network, redundant markers were deleted to construct a genetic linkage map;

[0054] Putting the phenotype data obtained by the MSN, ISN, and VSN methods into the composite interval mapping method (ICIM-ADD) of IciMapping 4.0 software, and combining the 660K SNP chip genotyping results to perform QTL mapping on the spike number per unit area trait of wheat. The results showed that the phenotypic results obtained based on this model had a relatively high correlation with the phenotypes obtained manually in the field and those based on picture annotation. A total of three QTL loci were co-located, among which the one on chromosome 7DS QSnyz.caas-7DS was located in all three ways, and the LOD value of this locus was between 3.34 and 4.86, and the physical interval was between 80.24 and 80.77, as Figure 5 shown. The above can indicate that the method in this embodiment can be used in the field of wheat breeding in a high-throughput manner.

[0055] In this embodiment, three QTL loci related to the spike number per unit area are located on wheat chromosomes 4DS, 7DS, and 7DL. The present invention can be used to screen wheat lines with a higher spike number per unit area, laying a theoretical foundation for breeding wheat varieties with high yield, stable yield, and excellent quality and providing a means of molecular assisted selection.

[0056] Example Two

[0057] Based on the same inventive concept, this embodiment discloses a wheat spike number recognition system based on Faster R-CNN, including:

[0058] A calibration module for extracting the canopy images of wheat ears of different families and calibrating the wheat ears in the images to obtain the labels corresponding to each wheat ear;

[0059] A feature map building module for inputting the images and labels of wheat ears into a ResNet network model for feature extraction to obtain the feature maps of wheat ears and establishing corresponding candidate boxes according to the outlines of wheat ears;

[0060] A model training module for training the Faster R-CNN model through the feature maps to obtain an optimal wheat ear recognition model;

[0061] A wheat ear recognition module for inputting the images of wheat ears to be measured into the wheat ear recognition model to obtain the candidate boxes corresponding to each wheat ear;

[0062] A wheat ear number output module for counting the number of candidate boxes to obtain the number of wheat ears.

[0063] Example Three

[0064] Based on the same inventive concept, this embodiment discloses a computer-readable storage medium storing one or more programs, where the one or more programs include instructions that, when executed by a computing device, cause the computing device to execute the Faster R-CNN-based wheat ear number recognition method according to any one of the above.

[0065] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0067] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention. The above content is only the specific embodiment of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or replacements, which should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A method for identifying the number of wheat ears based on Faster R-CNN, characterized in that, Including: Extracting the ear canopy images of different families, calibrating the ears in the images, and obtaining the labels corresponding to each ear; Inputting the images and labels of the ears into the ResNet network model for feature extraction, obtaining the feature maps of the ears, and establishing corresponding candidate boxes according to the ear contours; Training the Faster R-CNN model through the feature maps to obtain the optimal ear recognition model; Inputting the ear images to be measured into the ear recognition model to obtain the candidate boxes corresponding to each ear; Counting the number of the candidate boxes to obtain the number of wheat ears; Performing QTL mapping on the genetic traits of wheat through the number of ears per unit area; The method of the QTL mapping includes: Obtaining the doubled haploid population of wheat; Genotyping the doubled haploid population of wheat through an SNP chip; the genotyping method includes: performing whole-genome amplification on the genomic DNA of the wheat to be measured to obtain an amplification product; cutting the amplification product with a random endonuclease to obtain DNA fragments; hybridizing the DNA fragments with the SNP chip to make the DNA fragments complementary bind to the specific capture probes on the beads of the SNP chip; washing to remove the unhybridized or mismatched hybridized DNA fragments; performing single-base extension on the specific capture probes through dinitrophenol and biotin-labeled nucleotide substrates, and through staining, making different nucleotide substrates label different fluorescent dyes; scanning the SNP chip, judging and outputting the genotyping result according to the detected fluorescence number; inputting the genotyping result into the IciMapping network, deleting redundant markers, and constructing a genetic linkage map; Obtaining the QTL mapping of the wheat genetic traits through the composite interval mapping method according to the genetic linkage map, and mapping to three QTL loci related to the number of ears per unit area, which are distributed on wheat chromosomes 4DS, 7DS, and 7DL; The training of the Faster R-CNN model includes adjusting the ratio of the candidate boxes, the IOU threshold, and optimizing the strategy learning rate of the model training; Judging whether the candidate box exceeds the threshold through the L2 regularization of the fully connected layer and softmax classification and regression to judge whether the candidate box contains a complete ear. For the candidate box exceeding the threshold, obtaining its position coordinates through softmax regression, and removing the candidate box with an IOU threshold greater than that but not the optimal prediction of the real ear through non-maximum suppression; the images and labels of the ears are input into several sub-modules in the ResNet network model for feature extraction, generating candidate boxes according to the extracted features, performing L2 regularization and softmax classification and regression on the generated candidate boxes together with the original ear images and labels, calculating the intersection over union of the results, and performing non-maximum suppression on the results to obtain a preliminary classification scheme, and inputting the preliminary classification scheme into the ROI pooling layer, another sub-module in the ResNet network model, L2 regularization, softmax layer, intersection over union calculation, and non-maximum suppression in sequence to output the final image recognition result; Several sub-modules of the ResNet network model include four sub-modules. The first sub-module includes a convolutional layer, a BN layer, a RELU layer, and a max pooling layer. The second to fourth sub-modules each include a convolutional block and several recognition blocks. Another sub-module in the ResNet network model includes a convolutional block, two recognition blocks, a max pooling layer, and a flattening layer.

2. A wheat ear number recognition system based on Faster R-CNN, characterized in that, The wheat ear number recognition method based on Faster R-CNN according to claim 1 is adopted, including: A calibration module, configured to extract wheat ear canopy images of different families, calibrate the wheat ears in the images, and obtain labels corresponding to each wheat ear; A feature map building module, configured to input the images and labels of the wheat ears into the ResNet network model for feature extraction, obtain a feature map of the wheat ears, and build corresponding candidate boxes according to the wheat ear contours; A model training module, configured to train the Faster R-CNN model through the feature map to obtain an optimal wheat ear recognition model; A wheat ear recognition module, configured to input the image of the wheat ear to be measured into the wheat ear recognition model to obtain candidate boxes corresponding to each wheat ear; A wheat ear number output module, configured to count the number of the candidate boxes to obtain the number of wheat ears.

3. A computer-readable storage medium storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to execute the wheat ear number recognition method based on Faster R-CNN according to claim 1.

Images