Tassel recognition model training method and crop castration quality evaluation method and device

Abstract

The invention provides a tassel recognition model training method and a crop castration quality evaluation method and device.The tassel recognition model training method comprises the steps that crop images in a seed production field are screened and preprocessed, and a basic data set is obtained; labeling the tassels in various states in the basic data set according to the target features to obtain labeling boxes corresponding to the tassels in various states; updating all the annotation boxes in the basic data set according to the size and distribution characteristics of each annotation box to obtain an annotation data set; and performing iterative training on the convolutional neural network by taking the labeled data set as a training sample, taking cosine annealing learning rate scheduling as a learning rate strategy and taking the pre-training weight as an initial weight, and obtaining a crop tassel recognition model under the condition that the convolutional neural network converges or reaches the maximum number of iterations. According to the method, the identification precision of the non-outcrop tassels of the crops is improved, and then the corn castration quality evaluation efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a tassel recognition model training method, and a crop detasseling quality assessment method and device. Background Art

[0002] Hybrid crops (such as corn) need to be produced every year to ensure seed purity. During the corn seed production process, since the exposed male tassels often mean that pollen is about to begin to be shed, they must be removed as soon as possible to avoid adverse effects on seed purity after pollen shedding. Therefore, it is crucial to accurately identify the unemerged male tassels in advance, which not only provides time guarantee for the subsequent guidance of manual removal of the unemerged male tassels, but also effectively ensures the quality of emasculation.

[0003] In recent years, drone technology has been promoted and applied in all aspects of corn production, but its application in quality assessment after artificial emasculation is still in its infancy. Due to the short window period for corn emasculation, there is an urgent need to use drones to quickly and accurately detect the status of the female plant's male ears and evaluate the quality of artificial emasculation.

[0004] In the related technology, the existing technology uses feature engineering and deep learning methods to build a corn tassel recognition model that can easily identify the emerged tassels, but cannot identify the unemerged tassels. It can only continuously test and select a network structure suitable for corn tassel detection that meets the detection accuracy and efficiency, resulting in low efficiency in identifying tassels in different states and great difficulty in identifying unemerged tassels, resulting in low efficiency and accuracy in corn detasseling quality assessment. Summary of the invention

[0005] The present invention provides a tassel recognition model training method and a crop emasculation quality assessment method and device, which are used to solve the problem that the prior art is difficult to identify unemerged tassels using feature engineering and deep learning methods, and can only continuously test and select a network structure suitable for corn tassel detection and meeting detection accuracy and efficiency, resulting in low efficiency in identifying tassels in different states and great difficulty in identifying unemerged tassels, thereby improving the efficiency and accuracy of corn emasculation quality assessment.

[0006] The present invention provides a method for training a male tassel recognition model, comprising: Screen and preprocess the crop images in the seed production field to obtain the basic data set; Tassels of various states in the basic data set are labeled according to the target features to obtain labeling frames corresponding to the tassels of various states respectively; all labeling frames in the basic data set are updated according to the size and distribution characteristics of each labeling frame to obtain a labeling data set; wherein the target features include at least one of the difference between adjacent plants, leaf changes and color differences; the various states include fully exposed, partially exposed and not exposed; The labeled data set is used as a training sample, cosine annealing learning rate scheduling is used as a learning rate strategy, and pre-trained weights are used as initial weights to iteratively train the convolutional neural network. When the convolutional neural network converges or reaches a maximum number of iterations, a crop tassel recognition model is obtained.

[0007] According to a method for training a tassel recognition model provided by the present invention, after obtaining the labeled data set, the method further includes: Based on multiple detection models, model accuracy calculations are performed according to the sample annotation data set to obtain multiple model accuracy calculation results, and the detection model corresponding to the maximum value among the multiple model accuracy calculation results is determined as the benchmark model; The target size of the annotation box corresponding to the tassel in each state in the annotation data set is determined based on the benchmark model, and the sizes of all the annotation boxes in the annotation data set are adjusted according to the target size to obtain a new annotation data set.

[0008] According to a method for training a tassel recognition model provided by the present invention, the crop images in the seed production field are screened and preprocessed to obtain a basic data set including: Eliminating non-crop objects from the crop image to obtain a filtered crop image; At least one of image correction, image enhancement and cropping is performed on the screened crop images to obtain the basic data set.

[0009] According to a tassel recognition model training method provided by the present invention, the crop image is obtained by the following steps: The crop image is obtained by performing multi-temporal photography of the seed production field according to target parameters through unmanned aerial vehicle equipment; wherein the target parameters include camera parameters, flight parameters and environmental parameters; the flight parameters include flight altitude, speed and route, and the environmental parameters include light intensity, temperature and humidity.

[0010] According to a tassel recognition model training method provided by the present invention, after obtaining the crop tassel recognition model, the method further comprises: The crop tassel recognition model is optimized by at least one of the following methods: The annotated data set is expanded according to a target data enhancement strategy, and the crop tassel recognition model is optimized according to the expanded annotated data set to obtain an optimized crop tassel recognition model; wherein the target data enhancement strategy includes adjusting at least one of image brightness, contrast, hue, saturation and image shadow simulation; The crop tassel recognition model is adjusted according to the target update strategy to obtain an optimized crop tassel recognition model; wherein the target update strategy includes at least one of selecting a backbone network, adjusting network depth and width, feature fusion strategy, introducing an attention mechanism, and optimizing a loss function.

[0011] The present invention also provides a crop detasseling quality assessment method, comprising: Acquire a crop image to be identified; The crop image to be identified is processed based on the crop tassel identification model to obtain a tassel identification result; wherein the crop tassel identification model is trained by the tassel identification model training method; The quality of manual emasculation is evaluated according to the tassel identification result to obtain an evaluation result.

[0012] The present invention also provides a tassel recognition model training device, comprising: A data processing module is used to screen and preprocess crop images in the seed production field to obtain a basic data set; A data annotation module is used to annotate the tassels of various states in the basic data set according to the target feature, and obtain the annotation frames corresponding to the tassels of various states respectively; update all the annotation frames in the basic data set according to the size and distribution characteristics of each annotation frame, and obtain the annotation data set; wherein the target feature includes at least one of the difference between adjacent plants, leaf changes and color differences; and the various states include fully exposed, partially exposed and not exposed; The training module is used to iteratively train the convolutional neural network using the labeled data set as training samples, cosine annealing learning rate scheduling as the learning rate strategy, and pre-trained weights as initial weights, and obtain a crop tassel recognition model when the convolutional neural network converges or reaches a maximum number of iterations.

[0013] The present invention also provides a crop detasseling quality assessment device, comprising: An image acquisition module, used for acquiring images of crops to be identified; A recognition module, used for processing the crop image to be recognized based on a crop tassel recognition model to obtain a tassel recognition result; wherein the crop tassel recognition model is trained by the tassel recognition model training method; An evaluation module is used to evaluate the quality of manual emasculation according to the tassel identification result to obtain an evaluation result.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the tassel recognition model training method or crop detasseling quality assessment method as described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for training a male tassel recognition model or the method for assessing crop detasseling quality as described above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned methods for training a male tassel identification model or methods for assessing crop detasseling quality.

[0017] The tassel recognition model training method and crop tassel removal quality assessment method and device provided by the present invention mark tassels in various states in a basic data set by labeling features to obtain labeling frames corresponding to the tassels in various states respectively; update all labeling frames in the basic data set according to the size and distribution characteristics of each labeling frame to obtain a labeling data set; finally, the labeling data set is used as a training sample, cosine annealing learning rate scheduling is used as a learning rate strategy, and a convolutional neural network is iteratively trained with pre-training weights as initial weights to obtain a crop tassel recognition model, thereby improving the recognition accuracy of unexposed tassels of crops, and further improving the efficiency and accuracy of corn tassel removal quality assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a flow chart of the tassel recognition model training method provided by the present invention.

[0020] Figure 2 This is one of the flow charts of the crop detasseling quality assessment method provided by the present invention.

[0021] Figure 3 This is the second flow chart of the crop detasseling quality assessment method provided by the present invention.

[0022] Figure 4 It is a structural schematic diagram of the tassel recognition model training device provided by the present invention.

[0023] Figure 5 It is a structural schematic diagram of a crop detasseling quality assessment device provided by the present invention.

[0024] Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Combine the following Figure 1-Figure 5 The invention describes a male tassel recognition model training method and a crop detasseling quality assessment method and device.

[0027] Figure 1 : is a flow chart of the tassel recognition model training method provided by the present invention, such as Figure 1 As shown, the tassel recognition model training method includes the following steps: Step 110: Screen and preprocess the crop images in the seed production field to obtain a basic data set.

[0028] In this step, the crops in the seed field include but are not limited to one or more of corn, cotton, melons and fruits, etc.

[0029] In this step, the crop images can be obtained by taking low-altitude photos of the seed fields by aircraft such as drones, or can be directly obtained from a public database.

[0030] In this embodiment, the crop image quality is improved by screening out non-crop targets (field ridges, people, insects or other plants) in the crop image and retaining the crops to be identified.

[0031] In this embodiment, the basic data set is obtained by cropping, correcting, enhancing, etc. the crop images, thereby improving the analyzability of the images.

[0032] For example, by cropping the image, the high-resolution image is cropped into a batch of small images (such as squares with a side length of 1024), which are then sorted and classified. These small images will be used for annotation of subsequent detection models.

[0033] Step 120, labeling the tassels of various states in the basic data set according to the target features to obtain labeling boxes corresponding to the tassels of various states respectively; updating all the labeling boxes in the basic data set according to the size and distribution characteristics of each labeling box to obtain a labeling data set; wherein the target features include at least one of the differences between adjacent plants, leaf changes and color differences; and the various states include fully exposed, partially exposed and not exposed.

[0034] In this step, since the marking boundaries of the unemerged tassels of corn are not uniform, it is difficult to determine the size of the marking frame during the manual marking process; this embodiment uses the following steps to determine the optimal marking frame size: (1) Use a specified annotation box size to annotate all types of tassel targets in the small image.

[0035] In this embodiment, the size of the initial annotation box is usually determined based on expert experience and understanding of the morphology of corn plants; for example, professional image annotation tools (such as LabelImg or CVAT) can be used to perform preliminary annotations on tassels in different states.

[0036] It should be noted that in the preliminary labeling process, attention should be paid to the labeling of the unemerged tassels. The boundaries of the unemerged tassels can be predicted based on the differences between adjacent plants, slight changes in leaves and color differences, and then the labeling box of this type of tassel can be determined.

[0037] (2) Count the size and distribution (mean and variance) of the labeling box corresponding to each state of the tassel in the image to determine the approximate range of the labeling box.

[0038] Specifically, statistics are taken for different types of tassels (such as fully exposed, partially exposed, not exposed, etc.) to obtain the average width, height and standard deviation of each type of tassel annotation box, which provides basic data for subsequent optimization; the size adjustment interval and adjustment step of the annotation box are set to generate a series of annotation data sets, and a reasonable adjustment range (usually the mean value ± 2 standard deviations) is set based on the average width, height and standard deviation of each tassel annotation box. Within the adjustment range, multiple annotation boxes of different sizes are generated with a fixed step size (such as 5 or 10 pixels), thereby forming multiple versions of annotation data sets, i.e., annotation data sets.

[0039] Step 130: using the labeled data set as training samples, cosine annealing learning rate scheduling as the learning rate strategy, and pre-trained weights as initial weights to iteratively train the convolutional neural network, and obtaining a crop tassel recognition model when the convolutional neural network converges or reaches a maximum number of iterations.

[0040] In this step, the convolutional neural network includes but is not limited to the YOLO network, the RTMDet network, and the DINO network.

[0041] In this step, the convolutional neural network can be directly selected or screened out from multiple convolutional neural networks through model accuracy verification.

[0042] For example, multiple convolutional neural network models are trained using default parameters and weights, and then the average detection accuracy is used as an indicator to evaluate the performance of each model.

[0043] The following is an example of selecting a convolutional neural network to be trained from multiple convolutional neural networks.

[0044] In this embodiment, multiple convolutional neural networks are selected for evaluation; these models include but are not limited to the YOLO series (including YOLOv3, YOLOv4, YOLOv5, YOLOv7, etc.), Faster R-CNN and its variants, SSD (Single Shot Detector), RetinaNet, EfficientDet, DETR (DEtection TRansformer), and RTMDet (Real-time Object Detection), etc.

[0045] In this embodiment, the weights of each convolutional neural network are usually pre-trained on a large-scale dataset (such as COCO). Using pre-trained weights can speed up the convergence of the model and generally achieve better performance.

[0046] In this embodiment, when training a convolutional neural network using a labeled dataset, it can be divided into a training set, a validation set, and a test set according to a target ratio (e.g., 8:1:1) to ensure that the convolutional neural network has enough data for training while retaining enough data for verification and final performance evaluation.

[0047] In this embodiment, the following training strategies are used during the training of each model: (1) Learning rate strategy: cosine annealing learning rate scheduling is adopted; (2) Optimizer: Adam optimizer; (3) Batch size: adjusted according to GPU memory, usually 16 or 32; (4) Training rounds: fixed at 100 rounds, or use early stopping strategy.

[0048] After the training of each model is completed, the following evaluation indicators can be used to comprehensively evaluate the performance of each model: (1) mAP (mean Average Precision): calculated at different IoU thresholds (e.g. 0.5, 0.75); (2) Recall: evaluates the model detection rate; (3) Precision: evaluates the accuracy of the model; (4) F1 score: the harmonic mean of precision and recall; (5) FPS (Frames Per Second): evaluates the inference speed of the model.

[0049] In the process of selecting a suitable convolutional neural network and training a crop tassel recognition model, this implementation improves the performance and robustness of the crop tassel recognition model in a seed field environment by comprehensively considering the number of model parameters, detection efficiency, detection accuracy, and deployment application scenarios of each model.

[0050] In some embodiments, the convolutional neural network can also be determined by designing the following evaluation indicators: (6) Detection accuracy of unemerged tassels; (7) Performance under different lighting conditions, used to evaluate the stability of the model at different times and in different weather conditions; (8) Small target detection capability, used to evaluate the model’s ability to detect long-range or partially occluded tassels.

[0051] In this embodiment, the convolutional neural network to be trained is selected from multiple convolutional neural networks through the evaluation indicators corresponding to the above (1) to (8), and the labeled data set is used as the training sample, the cosine annealing learning rate scheduling is used as the learning rate strategy, and the pre-trained weight is used as the initial weight to iteratively train the convolutional neural network. When the convolutional neural network converges or reaches the maximum number of iterations, the crop tassel recognition model is obtained; this embodiment is aimed at the unexposed tassels that are difficult to identify in the seed production field, and the detection model with the highest detection accuracy is first selected. At the same time, considering the efficiency requirements in practical applications, the inference speed and resource consumption of the model are also weighed; in the model screening link, the present invention provides an artificially controllable model performance indicator selection strategy for obtaining the most suitable detection model in the current application scenario. This comprehensive evaluation strategy ensures that the selected model not only performs well in the laboratory environment, but also performs optimally in the actual corn seed production field.

[0052] The tassel recognition model training method provided by the embodiment of the present invention labels tassels of various states in a basic data set by labeling features to obtain labeling boxes corresponding to the tassels of various states respectively; updates all labeling boxes in the basic data set according to the size and distribution characteristics of each labeling box to obtain a labeling data set; finally, the labeling data set is used as a training sample, the cosine annealing learning rate scheduling is used as the learning rate strategy, and the pre-training weight is used as the initial weight to iteratively train the convolutional neural network to obtain a crop tassel recognition model, thereby improving the recognition accuracy of the unexposed tassels of the crop, and thereby improving the efficiency and accuracy of corn tassel removal quality assessment.

[0053] In some embodiments, after obtaining the labeled data set, the tassel recognition model training method also includes: performing model accuracy calculations based on multiple detection models according to the sample labeled data set, obtaining multiple model accuracy calculation results, and determining the detection model corresponding to the maximum value of the multiple model accuracy calculation results as the benchmark model; determining the target size of the labeling box corresponding to the tassel in each state in the labeled data set based on the benchmark model, and adjusting the size of all labeling boxes in the labeled data set according to the target size to obtain a new labeled data set.

[0054] In this embodiment, the method for acquiring the sample annotated data set is the same as the method for acquiring the annotated data set described above.

[0055] Specifically, the sample annotation datasets are used to train classic detection models (such as YOLO-v5, PP-YOLO, and PP-PicoDet networks) to obtain the detection accuracy of each model on the test set; by comparing the accuracies of different targets (such as the detection model with the highest accuracy), the optimal annotation size for each type of target is determined; after determining the annotation size for each type of target, the corresponding annotation box size in the dataset is modified to generate a new dataset for subsequent target detection model evaluation.

[0056] In this embodiment, YOLO-v5 (which performs well in target detection tasks and has a fast training speed) can be selected as the benchmark model. For each version of the data set, a complete model training and evaluation process is performed, and the detection accuracy of each category of targets is recorded. This embodiment not only focuses on the overall average accuracy, but also pays special attention to the detection accuracy of unexposed tassels. For each type of tassel, the annotation box size that can produce the highest detection accuracy is selected as the optimal size.

[0057] In this embodiment, part of the data is re-labeled using the YOLO-v5 model, or the size of the labeling box is automatically adjusted by a program.

[0058] The tassel recognition model training method provided by the embodiment of the present invention determines a baseline model by performing model accuracy calculation according to a sample annotation data set through multiple detection models, and determines the target size of the annotation box corresponding to the tassel in each state in the annotation data set according to the baseline model. Finally, the size of all the annotation boxes in the annotation data set is adjusted according to the target size, thereby realizing iterative optimization of the annotation boxes of tassels in different states to determine the annotation box size that is most suitable for each type of tassel, especially for the unexposed tassels that are difficult to identify, thereby improving the accuracy of subsequent target detection, and also providing a feasible annotation optimization method for other similar agricultural target detection tasks.

[0059] In some embodiments, crop images in a seed field are screened and preprocessed to obtain a basic data set, including: eliminating non-crop targets from the crop images to obtain screened crop images; performing at least one of image correction, image enhancement and cropping on the screened crop images to obtain a basic data set.

[0060] In this embodiment, since the UAV may have a slight tilt or shake during flight, causing a certain deformation of the captured image, the crop image is geometrically corrected by combining the UAV's posture data and ground control point information through image registration technology to ensure the spatial accuracy of the crop image.

[0061] In this embodiment, a histogram equalization operation is performed on the crop image to improve the overall brightness distribution of the image; an adaptive contrast enhancement operation is performed on the crop image to improve the visibility of local details; and a sharpening filter operation is performed on the crop image to enhance the edge details of the image. These image enhancement operations can improve the recognizability of the male spike in the crop image.

[0062] In this embodiment, image cropping is a key step in preprocessing; considering that the size of the original high-resolution image may reach thousands by thousands of pixels, directly using it for target detection will consume a lot of computing resources. In this embodiment, the original image is cropped into small images of a fixed size, usually 1024×1024 pixels, to ensure a balance between the integrity of a single corn plant and computing efficiency.

[0063] Specifically, in the process of cropping crop images, sliding window technology can be used to ensure that there is a certain overlap area between adjacent small images (usually 20%) to avoid the target being cropped and segmented; at the same time, the position information of each small image in the original crop image is also retained, which is crucial for subsequent result integration and spatial analysis.

[0064] In this embodiment, the cropped small images cover multiple types of tassel detection targets. For example, the small images include fully exposed mature tassels, tassels that are not exposed but are about to be pulled out, tassels that have been removed but have remaining parts, and the heart of the plant after the tassels have been removed. Through the above cropping method, the targets in the corn seed field can be analyzed more accurately, thereby providing a reliable basis for further decision-making.

[0065] The tassel recognition model training method provided in an embodiment of the present invention obtains a screened crop image by eliminating non-crop targets from the crop image; and performs at least one of image correction, image enhancement and cropping on the screened crop image, thereby improving the quality of the basic data set and providing high-quality input data for subsequent model training and evaluation.

[0066] In some embodiments, crop images are acquired through the following steps: multi-phase photography of a seed production field is performed by drone equipment according to target parameters to obtain crop images; wherein the target parameters include camera parameters, flight parameters and environmental parameters; flight parameters include flight altitude, speed and route, and environmental parameters include light intensity, temperature and humidity.

[0067] In this embodiment, a drone is used to capture a seed field scene of corn after manual emasculation during the tasseling period to obtain a crop image.

[0068] Specifically, advanced multi-rotor drone technology can be used to carry out low-altitude, high-definition photography of seed fields. During the shooting process, the camera parameters are carefully adjusted to ensure that the captured images have extremely high resolution, so that each corn plant and its individual tassel can be clearly identified.

[0069] For example, drones equipped with high-resolution cameras are used to capture crop images. The high-resolution cameras have a pixel count of more than 20 million and are equipped with an RTK positioning module to ensure the geographic positioning accuracy of the images.

[0070] In this embodiment, the flight altitude of the UAV equipment is usually set between 15-30 meters. This altitude range can increase the coverage area of ​​a single flight while ensuring image clarity; the flight speed of the UAV equipment is controlled at 3-5 meters per second to ensure that the image quality is not affected by motion blur; at the same time, the route planning is carried out for the entire seed production field, and the UAV equipment adopts a parallel line flight path to ensure that the selected areas in the seed production field can be effectively photographed.

[0071] In this embodiment, the shooting time of the drone equipment is selected in sunny weather with good lighting conditions to avoid the influence of cloudy days or strong light on the image quality. In order to capture the dynamic changes of corn growth, the present invention adopts a multi-phase shooting strategy, starting from the artificial emasculation of the seed production field, aerial photography is carried out every 1-2 days, and continues until the entire emasculation period; this high-frequency data collection can comprehensively record the changes of corn plants after artificial emasculation in the seed production field, and provide rich time series data for evaluating the quality of artificial emasculation.

[0072] The tassel recognition model training method provided in the embodiment of the present invention uses unmanned aerial vehicle equipment to perform multi-phase photography of seed fields according to camera parameters, flight parameters and environmental parameters, and can obtain a large amount of high-quality, multi-phase image data that meets the requirements of artificial emasculation in corn seed fields, thereby providing reliable data support for subsequent image processing and model training.

[0073] In some embodiments, after obtaining the crop tassel recognition model, the tassel recognition model training method further includes: optimizing the crop tassel recognition model by at least one of the following methods: First, the labeled data set is expanded according to the target data enhancement strategy, and the crop tassel recognition model is optimized according to the expanded labeled data set to obtain an optimized crop tassel recognition model; wherein the target data enhancement strategy includes adjusting at least one of image brightness, contrast, hue, saturation and image shadow simulation.

[0074] In this embodiment, the labeled data set is selectively expanded based on the fact that the field environment illumination is difficult to predict.

[0075] Specifically, in actual corn seed fields, lighting conditions may vary greatly due to factors such as weather, time, and occlusion. In order to enhance the model's ability to adapt to these changes, the following data augmentation strategies were implemented: (1) Brightness adjustment: randomly increase or decrease the overall brightness of the image to simulate lighting conditions at different time periods; (2) Contrast change: Adjust the contrast of the image to enhance the model’s recognition ability for low-contrast scenes.

[0076] (3) Hue shift: Slightly adjust the hue of the image to simulate color changes under different weather conditions.

[0077] (4) Saturation adjustment: Change the color saturation of the image to enhance the model’s ability to recognize corn at different growth stages.

[0078] (5) Shadow simulation: Add random shadows to the image to simulate clouds or other effects.

[0079] Through the above data enhancement strategy, the diversity of training data is effectively expanded. The retrained crop tassel recognition model can better adapt to various lighting and environmental conditions that may be encountered in actual farmland, improve the detection accuracy of the model, and enhance its stability under different weather and time conditions.

[0080] Secondly, the crop tassel recognition model is adjusted according to the target update strategy to obtain an optimized crop tassel recognition model; wherein the target update strategy includes at least one of selecting the backbone network, adjusting the network depth and width, feature fusion strategy, introducing the attention mechanism and optimizing the loss function.

[0081] In this embodiment, in order to further improve the detection accuracy of the crop tassel recognition model, the crop tassel recognition model can be optimized according to the following update strategy: (1) Backbone network selection: We tried a variety of mainstream convolutional neural networks as the backbone networks for feature extraction, including but not limited to the ResNet series, EfficientNet series, and MobileNet series. Through comparative experiments, we evaluated the performance of different networks in terms of accuracy, inference speed, and model size to find the backbone network that best suits the task of corn tassel detection.

[0082] (2) Network depth and width adjustment: For the selected backbone network, the network depth (number of layers) and width (number of channels per layer) are systematically adjusted to balance the expressiveness and computational complexity of the model; this process involves training and evaluating networks with different parameter scales to find the optimal network configuration.

[0083] (3) Feature fusion strategy: Considering that corn tassels may present features of different scales in an image, a multi-scale feature fusion mechanism is designed. This includes using a feature pyramid network (FPN) or its variants, and exploring weighted fusion strategies of features at different levels to improve the model’s ability to detect objects of different sizes.

[0084] (4) Introduction of attention mechanism: In order to enable the model to better focus on key areas and features, an attention mechanism was introduced into the network; this includes spatial attention and channel attention, which helps the model focus more on important information related to tassel detection during the feature extraction process.

[0085] (5) Loss function optimization: The loss function of the detection task was optimized, including the introduction of focal loss to solve the imbalance problem of positive and negative samples, and the design of special loss terms for small target detection to improve the model's sensitivity to difficult-to-detect targets such as unemerged male ears.

[0086] The tassel recognition model training method provided in the embodiment of the present invention effectively expands the diversity of training data through a variety of data enhancement strategies, enhances the stability and detection accuracy of the tassel recognition model under different weather and time conditions; and improves the detection accuracy and practicality of the tassel recognition model in processing complex backgrounds and small targets through the above-mentioned network architecture optimization strategy.

[0087] The crop detasseling quality assessment method provided by the present invention is described below. The crop detasseling quality assessment method described below and the tassel recognition model training method described above can be referenced to each other.

[0088] Figure 2 is one of the flow charts of the crop detasseling quality assessment method provided by the present invention, such as Figure 2 As shown, the crop detasseling quality assessment method comprises the following steps: Step 210: Acquire a crop image to be identified.

[0089] In this step, the image to be identified can be an image obtained from the above-mentioned annotated data set or public data set, or an image taken by a drone device in real time on a seed production field.

[0090] In this embodiment, after obtaining the crop image to be identified, the image segmentation strategy can be used to divide the crop image to be identified into multiple small images, and then the multiple small images are input into the subsequent crop tassel recognition model individually or in batches, so that the tassel detection results for the image or the entire plot can be quickly obtained, thereby improving the efficiency of crop tassel recognition.

[0091] It should be noted that since the crop tassel recognition model constructed subsequently can truly reflect the characteristics of the training data set, the selected model uses images of the same size as the training images when applied to reasoning; since the crop images to be identified are high-resolution drone images, they can be divided into dozens or hundreds of small images according to the size of the crop images to be identified, which will lead to two problems: the small images decomposed into a single image are all input into the model for reasoning, which significantly reduces the efficiency; when the results generated after the small image reasoning are fused, there are difficulties in fusing the edge targets in adjacent small images.

[0092] Specifically, this embodiment sets the overlapping area when performing image decomposition according to the size statistical information (average length, width, etc.) of the detection target, to ensure that each target to be detected can always be fully presented in a small image after decomposition, laying the foundation for subsequent target counting and segmentation. Furthermore, this embodiment calculates the distance between each drone image and the adjacent image based on the acquired batch image sequence of seed fields, thereby determining the effective utilization interval in each image; finally, based on the effective utilization interval, the minimum number of images that can be decomposed from the drone image can be calculated. Since the minimum number of images is much smaller than the image obtained by decomposing the entire crop image to be identified, the number of images used for reasoning can be greatly reduced.

[0093] In this embodiment, considering the efficiency requirements in practical applications, multiple blocks can be sent to different computing units (such as GPU) for detection at the same time through a parallel processing mechanism, which greatly improves the overall processing speed.

[0094] Through the above-mentioned image segmentation strategy, the detection problem of large-size images is converted into a parallel detection task of multiple small-size images, which not only greatly improves the detection efficiency, but also ensures the integrity and accuracy of the detection; experimental results show that compared with directly processing the original large-size image, the segmentation strategy reduces the processing time by more than 60% while maintaining similar detection accuracy.

[0095] Step 220: Process the crop image to be identified based on the crop tassel recognition model to obtain a tassel recognition result; wherein the crop tassel recognition model is trained by a tassel recognition model training method.

[0096] In this step, the tassel recognition model is trained through the following steps: (1) Screening and preprocessing crop images in seed production fields to obtain a basic data set; (2) labeling the tassels of various states in the basic data set according to the target features to obtain the labeling boxes corresponding to the tassels of various states respectively; updating all the labeling boxes in the basic data set according to the size and distribution characteristics of each labeling box to obtain the labeling data set; wherein the target features include at least one of the difference between adjacent plants, leaf changes and color differences; and the various states include fully exposed, partially exposed and not exposed; (3) The labeled dataset is used as training samples, the cosine annealing learning rate scheduling is used as the learning rate strategy, and the pre-trained weights are used as the initial weights to iteratively train the convolutional neural network. When the convolutional neural network converges or reaches the maximum number of iterations, a crop male spike recognition model is obtained.

[0097] In this embodiment, the training steps of the tassel recognition model are the same as those in the above-mentioned step 110 to step 130 corresponding to the embodiment, and will not be repeated in this embodiment.

[0098] Step 230: Perform manual detasseling quality assessment based on the tassel identification result to obtain an assessment result.

[0099] In this step, the tassel recognition result includes the number and coordinate information of exposed and unexposed tassels.

[0100] In this step, comprehensive manual detasseling quality assessment can provide quality feedback on the current detasseling work and also provide important guidance for subsequent precise detasseling operations.

[0101] The specific evaluation and application strategies are as follows: (1) Calculation of tassel density: First, calculate the density of exposed and uncovered tassels on a given plot. This calculation is based on the number of detected tassels and the actual area of ​​the plot; the density calculation formula is as follows: Density = number of detected tassels / plot area (m2); In this embodiment, the density of exposed tassels and the density of unexposed tassels are calculated respectively. These two indicators respectively reflect the effect of completed emasculation and the remaining workload to be processed.

[0102] (2) Evaluation of emasculation completion: Based on the ratio of exposed and unexposed tassels, the emasculation completion can be calculated using the following formula: Detasseling completion rate = number of exposed tassels / (number of exposed tassels + number of unexposed tassels) × 100%; In this embodiment, the degree of detasseling completion intuitively reflects the progress and effect of the current detasseling work.

[0103] (3) Spatial distribution analysis: Using the precise coordinate information of each tassel detected, a heat map of tassel distribution can be generated; this visualization method can intuitively show the spatial uniformity of emasculation work and help identify areas that may have been neglected or inadequately treated.

[0104] (4) Precise emasculation guidance: For each unemerged tassel detected, its precise location information (GPS coordinates) is recorded. This information can be directly output to agricultural machinery or handheld devices to guide workers to perform precise supplementary emasculation operations.

[0105] (5) Time series analysis: By comparing the test results of the same plot at different time points, the progress speed and efficiency of emasculation can be analyzed. This analysis helps to optimize human resource allocation and work plan formulation.

[0106] (6) Automatic report generation: Based on the above analysis results, the system can automatically generate a detailed quality assessment report; the report includes numerical statistics, chart visualization and specific improvement suggestions to provide decision support for managers.

[0107] The crop emasculation quality assessment method provided in the embodiment of the present invention processes the crop image to be identified through a crop tassel recognition model to obtain a tassel recognition result, thereby improving the tassel recognition efficiency and accuracy; the manual emasculation quality assessment method can timely and accurately evaluate the current emasculation effect, and can also provide detailed guidance for subsequent precise operations, thereby improving the efficiency and quality of the seed production process, and ultimately helping to improve the yield and quality of seed corn; at the same time, the data accumulation and analysis capabilities of this system provide data support for the intelligence and precision of future agricultural production.

[0108] Figure 3 This is the second flow chart of the crop detasseling quality assessment method provided by the present invention. Figure 3In the illustrated embodiment, in the training phase, UAV images of artificial emasculation seed production fields are first acquired through unmanned aerial vehicle equipment, and then different types of tassels are annotated on the image data through the established annotation frame optimization model to obtain an annotation data set; then a detection model evaluation strategy is established to select a convolutional neural network to be trained from multiple convolutional neural networks, and the convolutional neural network to be trained is trained using the annotated data set to obtain a tassel recognition model; the tassel recognition model is optimized through the model optimization strategy to obtain an optimized tassel recognition model; in the testing phase, an image segmentation strategy is established to segment the image to be recognized to obtain multiple small images, which are input into the optimized tassel recognition model for tassel recognition, and the tassel recognition results are used to evaluate the quality of artificial emasculation; finally, based on the above method, relevant products or devices are designed for use in artificial emasculation quality evaluation scenarios of crops such as corn.

[0109] The tassel recognition model training device provided by the present invention is described below. The tassel recognition model training device described below and the tassel recognition model training method described above can be referenced to each other.

[0110] Figure 4 : is a schematic diagram of the structure of the tassel recognition model training device provided by the present invention, such as Figure 4 As shown, the male spike recognition model training device includes: a data processing module 410, a data labeling module 420 and a training module 430.

[0111] The data processing module 410 is used to screen and pre-process the crop images in the seed production field to obtain a basic data set; The data annotation module 420 is used to annotate the tassels of various states in the basic data set according to the target features, and obtain the annotation frames corresponding to the tassels of various states respectively; update all the annotation frames in the basic data set according to the size and distribution characteristics of each annotation frame, and obtain the annotation data set; wherein the target features include at least one of the difference between adjacent plants, leaf changes and color differences; and the various states include fully exposed, partially exposed and not exposed; The training module 430 is used to iteratively train the convolutional neural network using the labeled data set as training samples, cosine annealing learning rate scheduling as the learning rate strategy, and pre-trained weights as initial weights, and obtain a crop tassel recognition model when the convolutional neural network converges or reaches the maximum number of iterations.

[0112] The tassel recognition model training device provided by the embodiment of the present invention labels tassels of various states in a basic data set by labeling features to obtain labeling boxes corresponding to the tassels of various states respectively; updates all labeling boxes in the basic data set according to the size and distribution characteristics of each labeling box to obtain a labeling data set; finally, the labeling data set is used as a training sample, the cosine annealing learning rate scheduling is used as the learning rate strategy, and the pre-training weight is used as the initial weight to iteratively train the convolutional neural network to obtain a crop tassel recognition model, thereby improving the recognition accuracy of the unexposed tassels of the crop, thereby improving the efficiency and accuracy of corn tassel removal quality assessment.

[0113] The crop detasseling quality assessment device provided by the present invention is described below. The crop detasseling quality assessment device described below and the crop detasseling quality assessment method described above can be referenced to each other.

[0114] Figure 5 Schematic diagram of the structure of the crop detasseling quality assessment device provided by the present invention. Figure 5 As shown, the crop detasseling quality assessment device includes: an image acquisition module 510, a recognition module 520 and an assessment module 530.

[0115] An image acquisition module 510 is used to acquire an image of a crop to be identified; The recognition module 520 is used to process the crop image to be recognized based on the crop tassel recognition model to obtain a tassel recognition result; wherein the crop tassel recognition model is trained by the tassel recognition model training method; The evaluation module 530 is used to evaluate the quality of manual emasculation according to the tassel identification result to obtain an evaluation result. Satellite communication intersatellite transmission signal system, signal system and signal processing method The crop emasculation quality assessment device provided in the embodiment of the present invention processes the crop image to be identified through the crop tassel recognition model to obtain the tassel recognition result, thereby improving the tassel recognition efficiency and accuracy; the manual emasculation quality assessment method can timely and accurately evaluate the current emasculation effect, and can also provide detailed guidance for subsequent precise operations, thereby improving the efficiency and quality of the seed production process, and ultimately helping to improve the yield and quality of seed corn; at the same time, the data accumulation and analysis capabilities of this system provide data support for the intelligent and precise agricultural production in the future.

[0116] Figure 6 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 6As shown, the electronic device may include: a processor (processor) 610 , a communication interface (Communications Interface) 620 , a memory (memory) 630 and a communication bus 640 , wherein the processor 610 , the communication interface 620 , and the memory 630 communicate with each other through the communication bus 640 . The processor 610 can call the logic instructions in the memory 630 to execute the male tassel recognition model training method, which includes: screening and preprocessing the crop images in the seed field to obtain a basic data set; marking the male tassels in various states in the basic data set according to the target features to obtain the marking boxes corresponding to the male tassels in various states; updating all the marking boxes in the basic data set according to the size and distribution characteristics of each marking box to obtain a marked data set; wherein the target features include at least one of the differences between adjacent plants, leaf changes and color differences; the multiple states include fully exposed, partially exposed and not exposed; using the marked data set as a training sample, using cosine annealing learning rate scheduling as a learning rate strategy, and using pre-trained weights as initial weights to iteratively train the convolutional neural network, and obtaining a crop male tassel recognition model when the convolutional neural network converges or reaches the maximum number of iterations.

[0117] Alternatively, the above-mentioned crop detasseling quality assessment method is executed, the method comprising: obtaining an image of a crop to be identified; processing the image of the crop to be identified based on a crop tassel recognition model to obtain a tassel recognition result; wherein the tassel recognition model is trained by a tassel recognition model training method; and performing manual detasseling quality assessment based on the tassel recognition result to obtain an assessment result.

[0118] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0119] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the male tassel recognition model training method provided by the above-mentioned methods, which method includes: screening and preprocessing crop images in a seed production field to obtain a basic data set; marking male tassels in various states in the basic data set according to target features to obtain marking boxes corresponding to the male tassels in various states; updating all marking boxes in the basic data set according to the size and distribution characteristics of each marking box to obtain a marked data set; wherein the target features include at least one of the differences between adjacent plants, leaf changes and color differences; the multiple states include fully exposed, partially exposed and not exposed; using the marked data set as a training sample, using cosine annealing learning rate scheduling as a learning rate strategy, and using pre-trained weights as initial weights to iteratively train the convolutional neural network, and obtaining a crop male tassel recognition model when the convolutional neural network converges or reaches the maximum number of iterations.

[0120] Alternatively, the above-mentioned crop detasseling quality assessment method is executed, the method comprising: obtaining an image of a crop to be identified; processing the image of the crop to be identified based on a crop tassel recognition model to obtain a tassel recognition result; wherein the tassel recognition model is trained by a tassel recognition model training method; and performing manual detasseling quality assessment based on the tassel recognition result to obtain an assessment result.

[0121] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the male tassel recognition model training method provided by the above-mentioned methods, the method comprising: screening and preprocessing crop images in a seed production field to obtain a basic data set; labeling male tassels of various states in the basic data set according to target features to obtain labeling boxes corresponding to the male tassels of various states; updating all labeling boxes in the basic data set according to the size and distribution characteristics of each labeling box to obtain a labeled data set; wherein the target features include at least one of the differences between adjacent plants, leaf changes and color differences; the multiple states include fully exposed, partially exposed and not exposed; using the labeled data set as a training sample, using cosine annealing learning rate scheduling as a learning rate strategy, and using pre-trained weights as initial weights to iteratively train the convolutional neural network, and obtaining a crop male tassel recognition model when the convolutional neural network converges or reaches the maximum number of iterations.

[0122] Alternatively, the above-mentioned crop detasseling quality assessment method is executed, the method comprising: obtaining an image of a crop to be identified; processing the image of the crop to be identified based on a crop tassel recognition model to obtain a tassel recognition result; wherein the tassel recognition model is trained by a tassel recognition model training method; and performing manual detasseling quality assessment based on the tassel recognition result to obtain an assessment result.

[0123] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0124] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0125] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a male spike recognition model, characterized in that: include: Screen and preprocess the crop images in the seed production field to obtain the basic data set; Annotating the tassels in various states in the basic data set according to the target features to obtain the annotation boxes corresponding to the tassels in various states respectively; updating all the annotation frames in the basic data set according to the size and distribution characteristics of each annotation frame to obtain an annotation data set; wherein the target feature includes at least one of the difference between adjacent plants, leaf changes and color differences; and the multiple states include fully exposed, partially exposed and not exposed; The labeled data set is used as a training sample, cosine annealing learning rate scheduling is used as a learning rate strategy, and pre-trained weights are used as initial weights to iteratively train the convolutional neural network. When the convolutional neural network converges or reaches a maximum number of iterations, a crop tassel recognition model is obtained.

2. The tassel recognition model training method according to claim 1, characterized in that: After obtaining the labeled data set, the method further includes: Based on multiple detection models, model accuracy calculations are performed according to the sample annotation data set to obtain multiple model accuracy calculation results, and the detection model corresponding to the maximum value among the multiple model accuracy calculation results is determined as the benchmark model; The target size of the annotation box corresponding to the tassel in each state in the annotation data set is determined based on the benchmark model, and the sizes of all the annotation boxes in the annotation data set are adjusted according to the target size to obtain a new annotation data set.

3. The tassel recognition model training method according to claim 1, characterized in that: The basic data set obtained by screening and preprocessing the crop images in the seed production field includes: Eliminating non-crop objects from the crop image to obtain a filtered crop image; At least one of image correction, image enhancement and cropping is performed on the screened crop images to obtain the basic data set.

4. The tassel recognition model training method according to claim 1, characterized in that: The crop image is obtained by the following steps: The crop image is obtained by performing multi-temporal photography of the seed production field according to target parameters through unmanned aerial vehicle equipment; wherein the target parameters include camera parameters, flight parameters and environmental parameters; the flight parameters include flight altitude, speed and route, and the environmental parameters include light intensity, temperature and humidity.

5. The tassel recognition model training method according to claim 1, characterized in that: After obtaining the crop tassel recognition model, the method further includes: The crop tassel recognition model is optimized by at least one of the following methods: The annotated data set is expanded according to a target data enhancement strategy, and the crop tassel recognition model is optimized according to the expanded annotated data set to obtain an optimized crop tassel recognition model; wherein the target data enhancement strategy includes adjusting at least one of image brightness, contrast, hue, saturation and image shadow simulation; The crop tassel recognition model is adjusted according to the target update strategy to obtain an optimized crop tassel recognition model; wherein the target update strategy includes at least one of selecting a backbone network, adjusting network depth and width, feature fusion strategy, introducing an attention mechanism, and optimizing a loss function.

6. A method for evaluating crop emasculation quality, characterized in that: include: Acquire a crop image to be identified; The crop image to be identified is processed based on a crop tassel identification model to obtain a tassel identification result; wherein the crop tassel identification model is obtained by training the tassel identification model training method according to any one of claims 1 to 5; The quality of manual emasculation is evaluated according to the tassel identification result to obtain an evaluation result.

7. A tassel recognition model training device, characterized in that: include: A data processing module is used to screen and preprocess crop images in the seed production field to obtain a basic data set; A data annotation module is used to annotate the tassels in various states in the basic data set according to target features, and obtain annotation boxes corresponding to the tassels in various states respectively; updating all the annotation frames in the basic data set according to the size and distribution characteristics of each annotation frame to obtain an annotation data set; wherein the target feature includes at least one of the difference between adjacent plants, leaf changes and color differences; and the multiple states include fully exposed, partially exposed and not exposed; The training module is used to iteratively train the convolutional neural network using the labeled data set as training samples, cosine annealing learning rate scheduling as the learning rate strategy, and pre-trained weights as initial weights, and obtain a crop tassel recognition model when the convolutional neural network converges or reaches a maximum number of iterations.

8. A crop detasseling quality assessment device, characterized in that: include: An image acquisition module, used for acquiring images of crops to be identified; A recognition module, used for processing the crop image to be recognized based on a crop tassel recognition model to obtain a tassel recognition result; wherein the crop tassel recognition model is obtained by training the tassel recognition model training method according to any one of claims 1 to 5; An evaluation module is used to evaluate the quality of manual emasculation according to the tassel identification result to obtain an evaluation result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images