Wheat flag leaf and wheat ear character intelligent identification method and system based on deep learning

Abstract

The invention relates to an intelligent identification method and system for wheat flag leaf and wheat ear characters based on deep learning. The method comprises the following steps: carrying out image acquisition on a wheat plant sample; preprocessing the wheat plant sample image, inputting the preprocessed wheat plant sample image into a pre-established target detection model, and estimating four characters of the wheat extraction degree, the wheat spikelet number, the wheat ear length and the wheat flag leaf length of the wheat plant sample; and based on the character estimation result of the wheat plant sample, performing stability identification on the four characters of the wheat plant sample, and visually displaying the identification result in an image form. The method can be widely applied to the technical field of crop monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of crop monitoring, and in particular relates to a method and system for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning. Background Art

[0002] Wheat is one of the most important food crops in the world, and its yield and quality are directly related to food security and economic benefits. In the process of wheat planting and promotion, it is extremely important to accurately and comprehensively identify the distinctiveness, consistency and stability of new varieties. Stability identification is not only a necessary step in variety protection, but also a key link in ensuring market access and sustainable agricultural development. Traditional wheat variety identification is cumbersome and labor-intensive, requiring long-term manual observation and recording of wheat's morphological characteristics and yield performance at different growth stages.

[0003] Existing identification methods have many shortcomings and mainly rely on on-site observation and judgment by experienced experts. Since these methods are sensitive to environmental conditions and easily affected by human factors, they lead to large errors. In addition, the process of data recording and analysis is very time-consuming and labor-intensive, and it is difficult to meet the needs of modern agricultural efficiency and intelligence. Traditional methods also require multiple sampling, which is somewhat destructive and is not conducive to continuous monitoring of large areas of farmland.

[0004] With the development of high-throughput phenotyping technology, the use of automated systems and machine learning models for wheat stability identification has become a trend. The application of such technologies can achieve non-destructive testing, multi-sample synchronous monitoring and real-time data collection, providing a reliable basis for accurately judging the characteristics of wheat varieties. In addition, efficient stability identification technology can also shorten the variety research and development cycle, improve breeding efficiency, and provide scientific support for the optimization of modern agricultural planting systems.

[0005] There are some devices for agricultural phenotyping on the market, but these devices can only simply record the apparent characteristics of wheat, but cannot achieve comprehensive stability characteristic monitoring and analysis, and it is difficult to provide results for variety differentiation and consistency judgment in a short time. Summary of the invention

[0006] In response to the above problems, the purpose of the present invention is to provide a method and system for intelligent identification of wheat flag leaf and ear traits based on deep learning. The collected wheat plant images are processed based on the deep learning method to obtain four types of traits including flag leaf length, extraction degree, ear length, and spikelet number. Finally, the measurement data is analyzed and visualized into image output to complete the stability identification.

[0007] To achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning, comprising the following steps: Image acquisition of wheat plant samples; The wheat plant sample images are preprocessed and then input into the pre-established target detection model to estimate the four traits of the wheat plant samples, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length; Based on the trait estimation results of the wheat plant samples, the stability of the four traits of the wheat plant samples was identified, and the identification results were visualized in the form of images.

[0008] Furthermore, the image acquisition of the wheat plant sample includes: Placing the treated wheat plant samples neatly on the sample placement platform, and ensuring that the ears of the wheat plant samples face upward; The images of the wheat plant samples on the sample placement platform are collected using an RGB camera and numbered and stored.

[0009] Furthermore, the sample placement platform adopts a black PVC board with grid markings, and the surface of the PVC board is provided with an anti-slip layer. When in use, the sample placement platform is placed on the anti-slip foot pad, and a tripod rotating assembly for adjusting the height of the sample placement platform is provided between the sample placement platform and the anti-slip foot pad.

[0010] Furthermore, the wheat plant sample image is preprocessed and then input into a pre-established target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length, including: Preprocessing the collected wheat plant sample images to obtain cropped images of a preset size as images to be detected; The image to be detected is input into a pre-established and trained target detection model to estimate four traits of wheat plant samples: wheat extraction degree, wheat spikelet number, wheat spike length, and wheat flag leaf length; Determine whether the trait estimation results meet the preset trait standards. If so, perform stability identification; otherwise, re-perform image preprocessing.

[0011] Furthermore, the image to be detected is input into a pre-established and trained target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length, including: Inputting the image to be detected into a pre-established target detection model, wherein the target detection model includes: a first point detection model, a second point detection model, a first frame detection model, and a second frame detection model; The positions of wheat spikelets are estimated based on the first point detection model, and the number of spikelet positions is used as the number of wheat spikelets; Based on the second point detection model, the positions of the wheat ear base and flag leaf pulvinus are estimated, and the wheat extraction degree is calculated; The positions of the wheat ear and flag leaf are estimated based on the first frame detection model and the second frame detection model, and the length of the wheat ear and the flag leaf are calculated.

[0012] Furthermore, the first point detection model and the second point detection model adopt a P2PNet model, and the first box detection model and the second box detection model adopt a yolo11n-obb model.

[0013] Further, the position of the wheat ear base and the flag leaf pulvinus is estimated based on the second point detection model, and the wheat extraction degree is calculated, including: The positions of the base of the wheat ear and the flag leaf pulvinus are estimated based on the second point detection model, and the point coordinate position information of the detection points of the base of the wheat ear and the flag leaf pulvinus in the image to be detected is obtained based on the estimated values; According to the point coordinate position information of the base of the wheat ear and the flag leaf pulvinus, the extraction degree length in pixel length units was calculated using the Euclidean distance formula, and then converted into the actual value of the wheat extraction degree through a conversion algorithm.

[0014] Further, the Euclidean distance formula is expressed as:

[0015] Where d represents the extraction degree; x represents the horizontal coordinate of the pixels at the base of the ear and the pulvinus of the flag leaf; y represents the vertical coordinate of the pixels at the base of the ear and the pulvinus of the flag leaf.

[0016] Further, the positions of the wheat ear and the flag leaf are estimated based on the first frame detection model and the second frame detection model respectively, and the length of the wheat ear and the length of the flag leaf are calculated, including: Estimate the positions of wheat ears and flag leaves based on the first frame detection model and the second frame detection model, and obtain frame coordinate position information of the wheat ear positioning frame and the flag leaf positioning frame in the image based on the estimated values; After the long sides of the wheat ear detection frame and the flag leaf detection frame are converted by a conversion algorithm, the actual values ​​of the wheat ear length and the wheat flag leaf length are obtained.

[0017] In a second aspect, the present invention provides an intelligent identification system for wheat flag leaf and wheat ear traits based on deep learning, comprising: An image acquisition module, used for acquiring images of wheat plant samples; The trait estimation module is used to pre-process the wheat plant sample image and input it into a pre-established target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length; The stability identification module is used to perform stability identification on four traits of wheat plant samples based on the trait estimation results of the wheat plant samples, and to visualize the identification results in the form of images.

[0018] The present invention adopts the above technical solution, which has the following advantages: 1. The present invention uses a deep learning method to extract features from wheat plant images, ensuring the consistency of data acquisition under different environmental conditions, improving the stability and reliability of monitoring, and the P2PNet model and the yolo11n-obb model have simple structures and strong portability. In actual production and life, the computing power of the equipment is not required, and the hardware equipment is easy to carry.

[0019] 2. The present invention adopts a deep learning model to analyze wheat plant traits, which can accurately extract multiple phenotypic parameters such as flag leaf length, extraction degree, ear length, and spikelet number. Compared with traditional manual measurement methods, it greatly improves the measurement efficiency and accuracy.

[0020] 3. The present invention utilizes a trained target detection model for stability identification. The target detection model is trained based on a large amount of labeled image data, which ensures the high accuracy and robustness of the target detection model, making the trait evaluation of wheat plants more scientific and objective.

[0021] Therefore, the present invention has broad application prospects in the technical fields of crop phenotype monitoring and stability identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Throughout the accompanying drawings, the same reference numerals are used to represent the same components. In the accompanying drawings: Figure 1 It is a flow chart of a method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning provided by an embodiment of the present invention; Figure 2 The invention provides cropping and enhancement of indoor wheat plant images; Figure 3 It is a P2PNet network structure provided by an embodiment of the present invention; Figures 4 (a) to 4 (f) are schematic diagrams of the rotatable detection box of the YOLO11N-OBB network provided in an embodiment of the present invention; Figure 4 (a) is a θ-based OBB (oriented bounding box); Figure 4 (b) is a Point-based OBB; Figure 4 (c) is an h-based OBB; Figure 4 (d) is a Mask OBB; Figure 4 (e) is l 1< l 2, at this time, the θ-based OBB is expressed as (cx, cy, h, w, θ), the Point-based OBB is expressed as (x1, y1, x2, y2, x3, y3, x4, y4), and the h-based OBB is expressed as (x1, y1, x2, y2, h); Figure 4 (f) is l 1> l 2, at this time, the θ-based OBB is expressed as (cx, cy, h, w, θ'), the Point-based OBB is expressed as (x4, y4, x1, y1, x2, y2, x3, y3), and the h-based OBB is expressed as (x4, y4, x1, y1, w), where (x1, y1), (x2, y2), (x3, y3), and (x4, y4) are the coordinates of the four vertices of the bounding box; Figures 5 (a) to 5 (d) are visualization output diagrams of stability identification provided by an embodiment of the present invention, wherein Figure 5 (a) is the extraction degree, Figure 5 (b) is the number of spikelets, Figure 5 (c) is the spike length identification, and Figure 5 (d) is the flag leaf length. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0024] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0025] In recent years, with the rapid development of deep learning technology, target detection methods have gradually become an important research direction in the field of computer vision. Among them, point detection and box detection are two basic target detection techniques. Point detection methods mainly focus on specific key points of the target in the image, such as feature points or center points, and are usually used in applications such as face recognition and action recognition; while box detection methods locate the target object by drawing a bounding box in the image, and are widely used in object recognition, autonomous driving, and security monitoring. However, in terms of wheat stability identification, related deep learning applications are still relatively scarce.

[0026] Therefore, in some embodiments of the present invention, a method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning is provided, the method comprising: collecting images of wheat plant samples; pre-processing the wheat plant sample images and inputting them into a pre-established target detection model, estimating the four traits of wheat extraction degree, wheat spikelet number, wheat ear length and wheat flag leaf length of the wheat plant samples; based on the trait estimation results of the wheat plant samples, the four traits of the wheat plant samples are stability identified, and the analysis results are visualized. The present invention realizes rapid and accurate analysis of wheat plant characteristics through a deep learning method, and has important practical application value. Such technology can not only improve the utilization efficiency of wheat germplasm resources, but also reduce dependence on artificial experience, making the stability identification process of wheat more scientific, accurate and efficient, and has broad application prospects in the field of crop phenotypic monitoring and stability identification technology, and can promote the intelligent development of modern agriculture.

[0027] Correspondingly, in other embodiments of the present invention, a deep learning-based intelligent identification system for wheat flag leaf and ear traits is provided.

[0028] Example 1 like Figure 1 As shown, the present invention provides a method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning, which comprises the following steps: 1) Collect images of wheat plant samples; 2) The wheat plant sample images are preprocessed and then input into the pre-established target detection model to estimate the four traits of the wheat plant samples: wheat extraction degree, wheat spikelet number, wheat spike length, and wheat flag leaf length; 3) Based on the trait estimation results of the wheat plant samples, the stability of the four traits of the wheat plant samples was identified, and the stability identification results were visualized in the form of images.

[0029] Furthermore, in the above step 1), when collecting images of wheat plant samples, the following steps are included: 1.1) Place multiple treated wheat plant samples neatly on the sample placement platform, and ensure that the ears of the wheat plant samples face upwards for easy photography; 1.2) Use a high-definition RGB camera to collect images of wheat plant samples on the sample placement platform, and number and store them.

[0030] Furthermore, in the above step 1.1), the sample placement platform uses a PVC board with grid markings, and the surface of the PVC board is provided with an anti-skid layer. The grid markings are used to achieve the conversion from pixel size to actual size, and the anti-skid layer is used to prevent the wheat plant samples from shifting during the placement of the wheat plant samples. Preferably, the PVC board is black, which can better highlight the structural characteristics of the wheat plant samples.

[0031] When in use, place the sample placement platform on the non-slip foot pad, and adjust the height of the non-slip foot pad through the tripod rotation assembly to ensure that it remains level and stable even on uneven ground; the non-slip foot pad is made of silicone material to ensure that the sample placement platform can both absorb the ground and not damage the laboratory or field surface.

[0032] Furthermore, the above step 2) includes the following steps: 2.1) Preprocessing the collected wheat plant sample images to obtain cropped images of a preset size as images to be detected; 2.2) Input the image to be detected into the pre-established and trained target detection model to estimate the four traits of wheat plant samples: wheat extraction degree, wheat spikelet number, wheat spike length, and wheat flag leaf length; 2.3) Determine whether the property estimation result meets the property standard according to the preset stability identification standard. If so, proceed to step 3) for stability identification. Otherwise, return to step 2.1) for re-image preprocessing.

[0033] The preset stability identification standard is set according to actual needs. If the value is obviously smaller than the value observed by naked eyes, such as the number of spikelets is 80-100 at a glance, but the predicted value is only 20-30, wheat ears may be missed during image cropping. At this time, the system will automatically mark images that differ from the stability identification standard by more than 20%, and these images need to be re-processed.

[0034] Furthermore, in the above step 2.1), if Figure 2 As shown, the collected wheat plant sample images are preprocessed to improve the prediction accuracy of the model. Specifically, the preprocessing method includes: Image cropping: Since the resolution of the original image is too large (6000×4000), in order to reduce the time spent in image annotation and model training and improve feasibility, this embodiment crops the image into multiple cropped images with a resolution of 1024×1024.

[0035] Furthermore, the above step 2.2) includes the following steps: 2.2.1) Input the image to be detected into a pre-established target detection model, wherein the target detection model includes a first point detection model, a second point detection model, a first frame detection model and a second frame detection model; the first point detection model is used to estimate the position of the spikelet, i.e., the number of spikelets; the second point detection model is used to estimate the position of the wheat ear base and the flag leaf pulvinus; the first frame detection model is used to estimate the length of the wheat ear; and the second frame detection model is used to estimate the length of the wheat flag leaf; 2.2.2) Estimate the positions of wheat spikelets based on the first point detection model, and use the number of spikelet positions as the number of wheat spikelets; In this embodiment, the output of the trained first point detection model is the position information of wheat spikelets. The present invention uses a wheat counting algorithm to calculate the number of target point position information output by the first point detection model as the final spikelet number; 2.2.3) Based on the second point detection model, the positions of the wheat ear base and flag leaf pulvinus are estimated, and the wheat extraction degree is calculated; 2.2.4) Based on the first frame detection model and the second frame detection model, the positions of the wheat ear and the flag leaf are estimated respectively, and the length of the wheat ear and the flag leaf are calculated.

[0036] Furthermore, in the above step 2.2.1), the training process of the target detection model includes the following steps: 2.2.1.1) Obtain a preset number of cropped images using the same method as step 2.1), enhance the cropped images, and annotate the datasets according to the four traits to obtain four types of training datasets; In this embodiment, when the cropped image is annotated, the following steps are included: First, use the Labelme point annotation tool to manually mark the corresponding coordinate point of each spikelet in the cropped wheat plant image, that is, use the horizontal and vertical coordinates to record whether the pixel point in the wheat plant image is a spikelet, and obtain the first training data set; Second, using the Labelme point annotation tool, the coordinate points of each spikelet base and flag leaf pulvinus of the cropped wheat plant image are annotated using the same method as the spikelet annotation to obtain the second training data set; Third, use the Labelme polygon annotation tool to manually annotate the position information of the four corners of the detection frame that closely fits the wheat ear for each wheat ear in the cropped wheat plant image, and obtain the third training data set; Fourth, use the Labelme polygon annotation tool to annotate the flag leaf of the cropped wheat plant image with the same method as annotating wheat ears to obtain the position information of the four corners of the detection frame, and obtain the fourth training data set.

[0037] 2.2.1.2) The pre-established first point detection model, second point detection model, first box detection model and second box detection model are trained using four types of training data sets respectively.

[0038] Preferably, in the above step 2.2.1.2), when training the established target detection model, first, the training data set is randomly divided into a training set, a validation set and a test set; then, after training with the training set, the model parameters with the best training effect are selected as the final target detection model parameters through the validation test results obtained from the validation set and the test set. By learning a large number of image samples of wheat plants in the training set, the target detection model can automatically learn the trait characteristics of each trait. Specifically, the wheat plant image to be detected is input into the target detection model, and the target detection model will classify the wheat plant image at the pixel level, and classify each pixel as belonging to a trait characteristic or background.

[0039] like Figure 3 As shown, in this embodiment, the first point detection model and the second point detection model both adopt the P2PNet model architecture, which is mainly composed of a feature extractor, a regression head and a classification head. Among them, the feature extractor part is based on the VGG16 network, which is used to extract features layer by layer of the input wheat plant image and generate multiple layers of feature maps with different resolutions. The feature extractor uses multiple convolutional layers to extract features layer by layer, and the ReLU activation function is applied after each convolutional layer to enhance the nonlinear expression ability and prevent overfitting, so as to better capture the detailed features in the wheat plant image; the regression head part is used for upsampling operation to restore the low-resolution wheat plant feature map, generate a higher-resolution feature map, and predict the specific coordinate position of the target point through the convolution layer; the classification head part is used to classify each candidate point to determine whether it is a positive sample. The predicted points and the real points are optimized through a 1-to-1 matching mechanism to ensure that the model can accurately locate the positions of spikelets, spike bases, and flag leaf pulvinus in wheat plants.

[0040] Specifically, the VGG16 network contains 16 weight layers (13 convolutional layers and 3 fully connected layers) and uses 3×3 small convolution kernels instead of large convolution kernels, which allows the network to capture more complex features of wheat plant images through multiple convolutions while reducing the number of parameters and improving nonlinear expression capabilities; the pooling layers all use 2×2 maximum pooling and downsampling, which gradually reduces the size of the feature map while retaining key information and reducing the amount of computation.

[0041] The regression head is divided into three stages. Each stage gradually restores the spatial resolution of the feature map through convolutional layers and upsampling layers, and generates high-precision coordinate information using the multi-scale features passed from the feature extractor.

[0042] The classification head is divided into three stages to generate category confidence for each predicted point, which is used to determine whether the point is the center of the target object. In each stage, the classification head outputs the classification probability to determine whether the point is the real target point (such as the center of an object) and uses the cross entropy threshold loss function for training.

[0043] As shown in Figure 4 (a) to Figure 4 (f), in this embodiment, the frame detection model adopts the yolo11n-obb model architecture, which is mainly composed of a backbone network (Backbone), a neck network (Neck) and a head network (Head). Among them, the backbone network is based on a deep convolutional neural network, which is used to extract features layer by layer from the input wheat plant image, generate multiple layers of feature maps with different resolutions, and use multiple convolutional layers to extract features layer by layer. After each convolutional layer, a ReLU activation function is applied to enhance the nonlinear expression ability and prevent overfitting, so as to better capture the detailed features of wheat ears and crown leaves; the neck network is used to fuse the feature maps and enhance the feature representation ability; the head network is responsible for outputting the location, category and direction information of the target.

[0044] Specifically, the convolution structure used in the backbone network is ResNet50, in which the convolution layer is usually designed with a small 3×3 convolution kernel, which enables the network to capture more complex features through multiple convolutions, while reducing the number of parameters and improving nonlinear expression capabilities; the maximum pooling layer uses a 2×2 window for downsampling, which effectively retains key feature information while reducing the size of the feature map and reducing computational complexity.

[0045] The neck network uses a feature pyramid network (FPN) to fuse feature maps and combine multi-scale features from different levels to enhance the multi-scale detection capabilities of wheat ears and crown leaves. This part gradually restores the spatial resolution of feature maps through multiple convolutional layers and upsampling layers, and generates high-quality target information in the feature maps, ensuring that the model can accurately identify targets regardless of their size.

[0046] The head network is responsible for outputting the detection results, including the location, category, and orientation information of the target. This part is usually divided into multiple stages, each of which generates outputs of different scales, and each output layer predicts the bounding box coordinates, category confidence, and orientation information of the target. At each stage, the head network outputs the category probability of the target and processes it through the Sigmoid activation function to determine whether each predicted box contains the target object, thereby optimizing the detection performance and efficiently and accurately detecting wheat ears and crown leaves. YOLOv11-OBB supports the detection of tilted bounding boxes, which is more suitable for wheat plants in natural growth state.

[0047] Furthermore, the above step 2.2.3) includes the following steps: 2.2.3.1) Estimate the positions of the base of the wheat ear and the flag leaf pulvinus based on the second point detection model, and obtain the point coordinate position information of the detection points of the base of the wheat ear and the flag leaf pulvinus in the image to be detected based on the estimated values; 2.2.3.2) The coordinate position information of the wheat ear base and flag leaf pulvinus is calculated using the Euclidean distance formula to obtain the extraction length in pixel length units, which is then converted into the actual value of the wheat extraction length using a conversion algorithm.

[0048] Specifically, the steps include: First, the extracted degree is converted from pixel length to actual length using the cropped image size and the scale of the grid markings in the image as a reference; Secondly, the wheat ear length measurement algorithm based on the matching of two sets of points is used to convert the extraction degree from pixel length to actual length.

[0049] In this example, the width of the wheat plant image is 1024 pixels, and each 1024 pixels includes 5 segments, each segment is 4.8 cm, so the scale factor in the algorithm used is 0.024 cm / pixel. When matching the coordinates of the ear base and the flag leaf pulvinus, the algorithm is designed to consider the pairing successful when the horizontal distance is less than 40 pixels, that is, the Euclidean distance between the two coordinate points is used as the extraction degree.

[0050] Specifically, the distance formula is expressed as:

[0051] In the formula, d represents the extraction degree; x represents the horizontal coordinate of the pixels at the base of the ear and the flag leaf pulvinus; y represents the vertical coordinate of the pixels at the base of the ear and the flag leaf pulvinus; the pixel length of the extraction degree can be calculated using this formula.

[0052] Furthermore, the above step 2.2.4) includes the following steps: 2.2.4.1) Estimate the positions of wheat ears and flag leaves based on the first frame detection model and the second frame detection model respectively, and obtain the frame coordinate position information of the wheat ear positioning frame and the flag leaf positioning frame in the image based on the estimated values.

[0053] In this embodiment, the first frame detection model after training outputs the coordinate information of the four corners of the wheat ear detection frame and the long side information of the detection frame; the present invention sorts the confidence scores generated after the reference image passes through the head network, and takes the target frames with the same number of wheat ears in the picture and scores from high to low as the final prediction frames; The trained second frame detection model outputs the flag leaf coordinate position information and the long side of the detection frame. The present invention sorts the confidence scores generated by the reference image after passing through the head network, and takes the target frames with the same number of flag leaves in the image and scores from high to low as the final prediction frames. 2.2.4.2) After converting the long sides of the wheat ear detection frame and the long sides of the flag leaf detection frame through a conversion algorithm, the actual values ​​of the wheat ear length and the wheat flag leaf length are obtained.

[0054] The long sides of the wheat ear detection frame and the flag leaf detection frame are converted by the conversion algorithm, which means that the long sides of the wheat ear detection frame and the flag leaf detection frame are converted from pixel length to actual length using the cropped image size and the scale in the grid markings in the image as a reference. Then, the actual values ​​of the wheat ear length and the wheat flag leaf length are obtained using the above formula.

[0055] The deep learning-based intelligent identification method for wheat flag leaf and ear traits provided by the present invention can accurately identify the specificity, consistency and stability of wheat varieties, and can also be used to quickly screen the key traits of new varieties. According to the analysis results of the identification model, the variety approval process can be optimized, and breeders and researchers can improve and screen varieties more efficiently, thereby promoting innovation in wheat breeding and improving yield and quality.

[0056] Example 2 This embodiment further introduces the method of the present invention by taking the processing of indoor wheat plant images as an example.

[0057] 1) Collect images of wheat plant samples.

[0058] Wheat plant sample images were collected, and a total of 80 indoor wheat plant sample images were obtained, including multiple different varieties and different growth stages. One image included 10 or more wheat plant samples.

[0059] 2) After preprocessing the collected wheat plant sample images, the pre-established target detection model is input to calculate the traits of the wheat plant samples, and different phenotypic parameters of the wheat plant samples are extracted based on the estimation results.

[0060] Preprocess the image: Figure 2 As shown, wheat plant samples are imaged to obtain the original wheat image; since the weight of the background pixels in the original image is too large, in order to enhance the model's ability to learn different features of the four wheat traits during target detection, the original image is cropped, and the image size after cropping is 1024×1024. In order to further enhance the generalization ability of the model, this embodiment also performs image enhancement processing on the cropped image.

[0061] Constructing datasets: Use the Labelme annotation tool to manually annotate the cropped images for different trait detection tasks and perform image enhancement. Therefore, it is necessary to train the model parameters separately and construct four types of datasets, including spikelet dataset, extraction degree dataset, spike length dataset, and flag leaf dataset.

[0062] When constructing the spikelet dataset, the wheat ear part of the original image needs to be completely cut out, and the spikelet part needs to be point-annotated using the Labelme annotation tool. A total of 100 annotated images were obtained, and the spikelet dataset was expanded to 660 images using image enhancement technology. The annotated images were randomly divided into the first training set, the first validation set, and the first test set in a ratio of approximately 7:2:1. That is, 462 images were used as training sets, 132 images were used as validation sets, and 66 images were used as test sets.

[0063] When constructing the extraction dataset, the wheat ear part of the original image needs to be completely cut out, and the base of the ear and the pulvinus of the flag leaf need to be marked using the Labelme annotation tool. A total of 100 annotated images were obtained, and the spikelet dataset was expanded to 492 images using image enhancement technology. The annotated images were randomly divided into the second training set, the second validation set, and the second test set in a ratio of approximately 7:2:1. That is, 344 images were used as training sets, 98 images were used as validation sets, and 50 images were used as test sets.

[0064] When constructing the ear length dataset, the ear part of the original image needs to be completely cut out, and the Labelme annotation tool is used to perform polygon annotation on the ear part. A total of 100 annotated images are obtained. The Yolo model automatically introduces data enhancement technology during the training process, so no manual enhancement is performed. The annotated images are randomly divided into the first training set, the first validation set, and the first test set in a ratio of approximately 7:2:1. That is, 70 images are used as training sets, 20 images are used as validation sets, and 10 images are used as test sets.

[0065] When constructing the flag leaf dataset, the flag leaf part of the original image needs to be completely cropped out, and the flag leaf part needs to be framed and labeled using the Labelme annotation tool. A total of 100 annotated images were obtained. The Yolo model automatically introduces data enhancement technology during the training process, so no manual enhancement is performed. The annotated images are randomly divided into the first training set, the first validation set, and the first test set in a ratio of approximately 7:2:1. That is, 70 images are used as training sets, 20 images are used as validation sets, and 10 images are used as test sets.

[0066] Object detection model training: ① Training environment This embodiment is based on the Ubuntu 20.04.6 LTS system, the GPU is NVIDIA GeForce RTX 4090, the CPU is Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz × 32, and the image processing language is Python3.9, pytorch2.4.1+cu118.

[0067] ② Training parameter setting In this embodiment, the target point detection model and the target frame detection model input the training samples from the training set into the network by batch training. The batch size of the training set and the validation set of the target point detection model is set to 8, that is, each batch includes 8 samples, 83 batches are required to traverse the spikelet data set once, and 31 batches are required to traverse the extraction degree data set once; the batch size of the training set and the validation set of the target frame detection model is set to 16, that is, each batch includes 16 samples. Since the training process adopts the dynamic enhancement technology of each round of training images, the batch required for traversing the data set once is not fixed. The model will learn the features of thousands of image variants in multiple rounds (epochs), that is, all training data are sent to the neural network once; each model sets the number of training network iterations to 100 rounds, the initial learning rate of the model is set to 0.001, and the others are set to default parameters. And the validation loss is monitored by setting the "early_stopping" module, that is, when the loss of the validation set decreases less than the learning rate of 0.001, the training stops, in order to prevent the model from overfitting.

[0068] Model evaluation: In order to verify whether the trait data estimated by the model can be used as a standard for stability identification, it is necessary to accurately measure the detection performance of the model. 2), Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and these three evaluation indicators are used to evaluate the relationship between the model prediction value and the true value, which are also important indicators that directly reflect the detection effect of the model.

[0069] ① Coefficient of determination (R 2 ) R 2 Measures the degree of linear fit between the model's predicted results and the actual results. The calculation formula is as follows:

[0070] ② Mean absolute error (MAE) MAE measures the average absolute error between the predicted value and the true value. The calculation formula is as follows:

[0071] ③Root mean square error (RMSE) RMSE is used to measure the difference between the predicted value and the true value, reflecting the standard deviation of the predicted value. The calculation formula is as follows:

[0072] Among them, RSS refers to the residual sum of squares, that is, the sum of squares of the difference between the true value of each wheat trait and the model predicted value; TSS refers to the overall sum of squares, that is, the sum of squares of the difference between the true value of each wheat trait and the true average value.

[0073] The present invention uses the target detection models trained by P2PNet and yolo11n-obb networks to detect the first test set, the second test set, the third test set, and the fourth test set images, and calculates three evaluation indicators, taking the average value of each indicator as a reference indicator. The detection results are shown in the table.

[0074]

[0075] Based on the verification results of this table and Figure 5a ~ Figure 5d The stability identification output of this embodiment shown indicates that the estimation of various traits by the target detection model used in the present invention basically meets the stability identification requirements.

[0076] Example 3 The above-mentioned embodiment 1 provides an intelligent identification method for wheat flag leaves and wheat ears based on deep learning. Correspondingly, this embodiment provides an intelligent identification system for wheat flag leaves and wheat ears based on deep learning. The system provided in this embodiment can implement the intelligent identification method for wheat flag leaves and wheat ears based on deep learning in embodiment 1, and the system can be implemented by software, hardware, or a combination of software and hardware. For example, the system may include integrated or separate functional modules or functional units to execute the corresponding steps in each method of embodiment 1. Since the system of this embodiment is basically similar to the method embodiment, the process described in this embodiment is relatively simple, and the relevant parts can refer to the partial description of embodiment 1. The embodiment of the system provided in this embodiment is only schematic.

[0077] The deep learning-based intelligent identification system for wheat flag leaf and wheat ear traits provided in this embodiment includes: An image acquisition module, used for acquiring images of wheat plant samples; The trait estimation module is used to pre-process the wheat plant sample image and input it into a pre-established target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length; The stability identification module is used to perform stability identification on four traits of wheat plant samples based on the trait estimation results of the wheat plant samples, and to visualize the identification results in the form of images.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning, characterized in that: The following steps are involved: Image acquisition of wheat plant samples; The wheat plant sample images are preprocessed and then input into the pre-established target detection model to estimate the four traits of the wheat plant samples, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length; Based on the trait estimation results of the wheat plant samples, the stability of the four traits of the wheat plant samples was identified, and the identification results were visualized in the form of images.

2. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 1, characterized in that: The step of collecting images of wheat plant samples comprises: Placing the treated wheat plant samples neatly on the sample placement platform, and ensuring that the ears of the wheat plant samples face upward; The images of the wheat plant samples on the sample placement platform are collected using an RGB camera and numbered and stored.

3. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 2, characterized in that: The sample placement platform is made of a black PVC board with grid markings, and the surface of the PVC board is provided with an anti-skid layer. When in use, the sample placement platform is placed on the anti-skid foot pad, and a tripod rotating assembly for adjusting the height of the sample placement platform is provided between the sample placement platform and the anti-skid foot pad.

4. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 1, characterized in that: The wheat plant sample image is pre-processed and then input into a pre-established target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length, including: Preprocessing the collected wheat plant sample images to obtain cropped images of a preset size as images to be detected; The image to be detected is input into a pre-established and trained target detection model to estimate four traits of wheat plant samples: wheat extraction degree, wheat spikelet number, wheat spike length, and wheat flag leaf length; Determine whether the trait estimation results meet the preset trait standards. If so, perform stability identification; otherwise, re-perform image preprocessing.

5. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 4, characterized in that: The image to be detected is input into a pre-established and trained target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length, including: Inputting the image to be detected into a pre-established target detection model, wherein the target detection model includes: a first point detection model, a second point detection model, a first frame detection model, and a second frame detection model; The positions of wheat spikelets are estimated based on the first point detection model, and the number of spikelet positions is used as the number of wheat spikelets; Based on the second point detection model, the positions of the wheat ear base and flag leaf pulvinus are estimated, and the wheat extraction degree is calculated; The positions of the wheat ear and flag leaf are estimated based on the first frame detection model and the second frame detection model, and the length of the wheat ear and the flag leaf are calculated.

6. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 5, characterized in that: The first point detection model and the second point detection model adopt the P2PNet model, and the first box detection model and the second box detection model adopt the yolo11n-obb model.

7. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 5, characterized in that: The method of estimating the position of the wheat ear base and the flag leaf pulvinus based on the second point detection model and calculating the wheat extraction degree includes: The positions of the base of the wheat ear and the flag leaf pulvinus are estimated based on the second point detection model, and the point coordinate position information of the detection points of the base of the wheat ear and the flag leaf pulvinus in the image to be detected is obtained based on the estimated values; According to the point coordinate position information of the base of the wheat ear and the flag leaf pulvinus, the extraction degree length in pixel length units was calculated using the Euclidean distance formula, and then converted into the actual value of the wheat extraction degree through a conversion algorithm.

8. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 7, characterized in that: The Euclidean distance formula is expressed as: ； Where d represents the extraction degree; x represents the horizontal coordinate of the pixels at the base of the ear and the pulvinus of the flag leaf; y represents the vertical coordinate of the pixels at the base of the ear and the pulvinus of the flag leaf.

9. The method for intelligent identification of wheat flag leaf and wheat ear traits based on deep learning as claimed in claim 5, characterized in that: The method of estimating the positions of wheat ears and flag leaves based on the first frame detection model and the second frame detection model, and calculating the length of the wheat ears and the length of the flag leaves, comprises: Estimate the positions of wheat ears and flag leaves based on the first frame detection model and the second frame detection model, and obtain frame coordinate position information of the wheat ear positioning frame and the flag leaf positioning frame in the image based on the estimated values; After the long sides of the wheat ear detection frame and the flag leaf detection frame are converted by a conversion algorithm, the actual values ​​of the wheat ear length and the wheat flag leaf length are obtained.

10. A deep learning-based intelligent identification system for wheat flag leaf and ear traits, characterized in that: include: An image acquisition module, used for acquiring images of wheat plant samples; The trait estimation module is used to pre-process the wheat plant sample image and input it into a pre-established target detection model to estimate four traits of the wheat plant sample, namely, wheat extraction degree, wheat spikelet number, wheat spike length and wheat flag leaf length; The stability identification module is used to perform stability identification on four traits of wheat plant samples based on the trait estimation results of the wheat plant samples, and to visualize the identification results in the form of images.

Images