A field corn seed real-time identification method based on an edge computing platform and an electronic device

Abstract

This invention proposes a real-time field maize seed identification method and electronic device based on an edge computing platform. It solves the problems of high false negative rates for small maize seeds, numerous false positives in complex backgrounds, and the difficulty of low-latency real-time operation of deep learning models on embedded platforms in existing technologies. This invention collects field images with varying lighting conditions, soil types, and operating speeds, performs quality screening and cross-labeling to construct a highly diverse standardized dataset. An improved YOLO model incorporating feature pyramids, attention mechanisms, and anchor box redesign is used for training. The trained PyTorch model is exported sequentially to ONNX format, and an FP16 precision inference engine is built based on TensorRT. Finally, it is deployed to an embedded edge computing platform, achieving real-time acquisition, inference, and result output at a rate of no less than 30 frames per second through a pipelined parallel architecture. This invention significantly reduces inference latency and power consumption while maintaining high detection accuracy and can be integrated into intelligent agricultural machinery vision inspection and agricultural edge computing systems.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, artificial intelligence and embedded edge computing, and in particular to a method and electronic device for real-time identification of corn seeds in the field based on an edge computing platform. Background Technology

[0002] In modern agricultural applications such as precision corn planting, seed quality testing, and field emergence monitoring, achieving automatic, rapid, and accurate identification of corn seeds in the field is crucial for improving the level of intelligent operations. Early solutions relied on manual visual inspection. This method, entirely dependent on operators observing, counting, or judging by eye, was inefficient, labor-intensive, and its results were heavily influenced by personal experience and subjective states, failing to meet the demands of large-scale, automated agricultural production. With the development of computer vision technology, methods based on traditional image processing have emerged. These methods typically employ threshold segmentation, edge detection, morphological operations, and strategies combining color and shape-based feature extraction with classifiers. For example, published literature includes schemes that segment the soil background using the Otsu thresholding method and then use Hough transform to detect circular contours for seed identification. However, field environments are characterized by uneven lighting, complex and variable soil backgrounds, and seeds being partially occluded or similar in color to stubble and debris. This leads to insufficient robustness of features designed by traditional algorithms, poor generalization ability under changing conditions, and high false positive and false negative rates.

[0003] In recent years, deep learning-based object detection algorithms have become mainstream, especially models such as the YOLO series, Faster R-CNN, and SSD, which have demonstrated superior performance in general object detection tasks. Researchers have begun to try applying these general models directly to agricultural detection tasks. However, when directly transferring such general models to the specific scenario of corn seed recognition in the field, three bottlenecks are encountered. First, corn seeds in field images exhibit typical small target characteristics, and the high-level feature maps of general detection models do not retain enough semantic information of small targets, resulting in low detection accuracy and serious missed detections. Second, complex field background interference (such as soil particles, weeds, and shadows) leads to a high false detection rate. Third, these models usually have a large number of parameters and high computational complexity. If directly deployed on embedded edge computing platforms with limited computing power, memory, and power consumption, it is difficult to meet the requirements of real-time inference and low-power operation. Although some studies have attempted to lightweight the models through model pruning and quantization, they often lack systematic solutions for balancing accuracy and speed, as well as for optimization for specific scenarios.

[0004] In summary, existing technologies cannot simultaneously achieve high-precision, low-latency, and robust real-time recognition of small corn seeds in complex field environments on embedded edge computing platforms with limited computing power. Traditional image processing methods have poor environmental adaptability; general-purpose deep learning models perform poorly in small target detection and are difficult to meet the real-time and low-power requirements of embedded deployments. Summary of the Invention

[0005] To address the challenges of small targets, complex backgrounds, and embedded real-time deployment in field maize seed identification, this invention provides a real-time field maize seed identification method based on an edge computing platform. By improving the target detection algorithm structure, data acquisition and annotation strategies, and combining an embedded edge computing platform with TensorRT inference acceleration technology, this method achieves high-precision, low-latency, and low-power real-time identification of field maize seeds. The method includes: Step 1: Collect raw image data containing corn seeds under different field conditions; Step 2: Perform data preprocessing and quality screening on the original image data to obtain a set of qualified images; Step 3: Label and perform quality control on the qualified image set to generate a standardized dataset containing labeled bounding boxes; Step 4: Perform data augmentation and partitioning on the standardized dataset to construct training, validation, and test sets; Step 5: Use the training set, validation set, and test set to train the YOLO model to obtain the trained model; Step 6: Export the trained model as an ONNX format model; Step 7: Optimize the ONNX format model based on TensorRT to generate a TensorRT inference engine file suitable for embedded platforms; Step 8: Deploy the TensorRT inference engine file to an embedded edge computing platform to perform real-time image acquisition, image preprocessing, TensorRT accelerated inference, detection result post-processing, and recognition result output.

[0006] Furthermore, in step one, The field environment includes different lighting conditions, different soil types, and different operating speeds; The image resolution used during acquisition ranges from 1280×720 to 1920×1080, the shooting height ranges from 20cm to 80cm, and the shooting frame rate ranges from 15fps to 60fps.

[0007] Furthermore, in step two, Data preprocessing and quality screening include: sharpness detection of the original image data to remove blurry images; exposure detection of the original image data to remove severely underexposed or overexposed images; and deduplication of duplicate and highly similar images from the original image data.

[0008] Furthermore, in step three, Labeling and quality control include: labeling bounding boxes and assigning category labels to corn seed targets in qualified image sets; performing multi-person cross-labeling for difficult samples; and sampling review and consistency verification of the labeling results. The minimum labeled target size is 10×10 pixels, and the allowed labeled proportion for occluded targets is ≥30% of the visible area.

[0009] Furthermore, in step four, Data augmentation of standardized datasets includes: brightness perturbation, contrast perturbation, color perturbation, random rotation, random scaling, random cropping, and adding noise to simulate field disturbances.

[0010] Furthermore, in step five, The YOLO model includes one or more of the following improvements: introducing a feature pyramid structure suitable for small object detection, introducing an attention mechanism, redesigning anchor box parameters, and adjusting the weight ratio of classification loss to localization loss. The YOLO model input image size is 1920×1080; the confidence threshold Tc is 0.3 to 0.7; and the non-maximum suppression (NMS) threshold Tn is 0.3 to 0.6.

[0011] Furthermore, step six specifically involves: Perform model structure freezing; export ONNX intermediate representation; perform operator compatibility checks and graph structure optimization on the ONNX graph.

[0012] Furthermore, step seven specifically involves: FP16 was selected as the inference precision mode; execution layer fusion and memory optimization were implemented; and TensorRT inference engine files suitable for embedded platforms were generated.

[0013] Furthermore, in step eight, The embedded edge computing platform is the NVIDIA Jetson AGX Orin platform; after deployment, the real-time inference speed is no less than 30 frames per second.

[0014] The present invention proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0015] The beneficial effects of this invention are: This invention addresses the challenges of small targets, complex backgrounds, and embedded real-time deployment in field maize seed identification. It provides a method and electronic device for real-time field maize seed identification based on an edge computing platform. By improving the target detection algorithm structure, data acquisition and annotation strategies, and combining an embedded edge computing platform with TensorRT inference acceleration technology, it achieves high-precision, low-latency, and low-power real-time identification of field maize seeds, with the following improvements: 1. This invention establishes a multi-scenario field image acquisition standard covering different lighting conditions, soil types, and operating speeds, and implements strict quality screening and annotation review to construct a high-quality, highly diverse standardized dataset, which significantly improves the model's generalization ability and robustness in real and complex field environments.

[0016] 2. This invention significantly enhances the feature extraction and expression capabilities for small-sized corn seeds by introducing at least one of the following optimization structures into the YOLO model: a feature pyramid structure suitable for small object detection, an attention mechanism, redesigned anchor box parameters, and adjusted weight ratios of the loss function. This also greatly improves the recall rate of small objects while maintaining background suppression capabilities.

[0017] 3. This invention constructs a complete model conversion pipeline encompassing PyTorch training, ONNX export, and TensorRT optimization. It also employs FP16 precision inference, layer fusion, and memory optimization strategies to generate an inference engine highly adapted to embedded platform hardware. This significantly reduces model computational complexity and GPU memory usage with almost no loss in detection accuracy.

[0018] 4. This invention deploys the optimized TensorRT inference engine on an embedded edge computing platform and realizes parallel processing of the entire pipeline from image acquisition, preprocessing, accelerated inference and result output. It achieves a real-time inference speed of no less than 30 frames per second on platforms such as NVIDIA Jetson AGXOrin, providing low-latency visual feedback and control interfaces for intelligent agricultural machinery.

[0019] 5. Through complete algorithm optimization and system-level architecture design, this invention achieves low-power, low-latency real-time recognition capabilities without relying on high-performance computing devices, significantly reducing the operating power consumption and heat dissipation requirements of embedded platforms, and greatly enhancing the engineering and industrial application value of this invention in the field of agricultural intelligent equipment. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention proposes a real-time field corn seed identification method based on an edge computing platform, the method comprising: Step 1: Collect raw image data containing corn seeds under different field conditions; The field environment includes different lighting conditions, different soil types, and different operating speeds; The image resolution used during acquisition ranges from 1280×720 to 1920×1080, the shooting height ranges from 20cm to 80cm, and the shooting frame rate ranges from 15fps to 60fps.

[0022] Data was collected using image acquisition modules deployed on simulated or real agricultural machinery, with resolutions ranging from 1280×720 pixels to 1920×1080 pixels. During acquisition, cameras were mounted at a height of 20 cm to 80 cm above the ground, capturing images at a fixed or adjustable frame rate (15 fps to 60 fps). To ensure the diversity and robustness of the dataset, a large number of images containing maize seeds were collected at different times, under different weather conditions, with different soil types, and under different field operation conditions. The collected raw image data was stored on storage media for subsequent processing.

[0023] Step 2: Perform data preprocessing and quality screening on the original image data to obtain a set of qualified images; Data preprocessing and quality screening include: sharpness detection of the original image data to remove blurry images; exposure detection of the original image data to remove severely underexposed or overexposed images; and deduplication of duplicate and highly similar images from the original image data.

[0024] The obtained raw image data is cleaned to remove low-quality samples and improve the overall effectiveness of the dataset.

[0025] Step 3: Label and perform quality control on the qualified image set to generate a standardized dataset containing labeled bounding boxes; Labeling and quality control include: labeling bounding boxes and assigning category labels to corn seed targets in qualified image sets; performing multi-person cross-labeling for difficult samples; and sampling review and consistency verification of the labeling results. The minimum labeled target size is 10×10 pixels, and the allowed labeled proportion for occluded targets is ≥30% of the visible area.

[0026] The labeling work was carried out using professional labeling software. Labellers used rectangular boxes to label each visible corn seed according to specifications, and assigned a category label "corn seed" to each box. For particularly difficult cases, such as severely obscured seeds or seeds that were extremely similar in color to the background, a multi-person cross-labeling method was used to ensure accuracy.

[0027] Step 4: Perform data augmentation and partitioning on the standardized dataset to construct training, validation, and test sets; Data augmentation of standardized datasets includes: brightness perturbation, contrast perturbation, color perturbation, random rotation, random scaling, random cropping, and adding noise to simulate field disturbances.

[0028] Step 5: Using the training set, validation set, and test set, train the improved YOLO model to obtain the trained model; The improved YOLO model includes the following optimizations: introducing a feature pyramid structure suitable for small object detection, introducing an attention mechanism, redesigning anchor box parameters, and adjusting the weight ratio of classification loss to localization loss. The improved YOLO model has an input image size of 1920×1080; a confidence threshold Tc of 0.3–0.7; and an NMS threshold Tn of 0.3–0.6.

[0029] Using a pre-constructed dataset, a deep learning model specifically designed for small target detection in maize seeds in the field was trained. Training was conducted using the PyTorch deep learning framework. Based on the YOLO network structure, the model underwent the following targeted improvements: A feature pyramid structure was introduced or strengthened into the network to better integrate deep semantic features with shallow detail features, enhancing small target detection capabilities; an attention mechanism module was introduced into the key feature layer to make the model focus more on the seed region and suppress interference from complex backgrounds; the size and proportion of the default anchor boxes were redesigned based on the statistical clustering results of the labeled boxes in the training set to better reflect the actual size of maize seeds; and the weight ratio of classification loss to bounding box localization loss in the loss function was adjusted to optimize the model's performance in accurately locating small targets. During training, input images were uniformly scaled to a fixed size (e.g., 640×640 pixels), and the model weights that performed best on the test set were saved.

[0030] Step 6: Export the trained model as an ONNX format model; Perform model structure freezing; export ONNX intermediate representation; perform operator compatibility checks and graph structure optimization on the ONNX graph.

[0031] Step 7: Optimize the ONNX format model based on TensorRT to generate a TensorRT inference engine file suitable for embedded platforms; FP16 was selected as the inference precision mode; execution layer fusion and memory optimization were implemented; and TensorRT inference engine files suitable for embedded platforms were generated.

[0032] Step 8: Deploy the TensorRT inference engine file to an embedded edge computing platform to perform real-time image acquisition, image preprocessing, TensorRT accelerated inference, detection result post-processing, and recognition result output.

[0033] The embedded edge computing platform is the NVIDIA Jetson AGX Orin platform; after deployment, the real-time inference speed is no less than 30 frames per second.

[0034] The present invention proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0035] The above provides a detailed description of the real-time identification method and electronic device for maize seeds in the field based on an edge computing platform proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for real-time identification of corn seeds in the field based on an edge computing platform, characterized in that, The method comprises: Step 1: Collecting original image data containing corn seeds under different field environments; Step 2: Preprocessing and quality screening of the original image data to obtain a qualified image set; Step 3: Labeling and quality control of the qualified image set to generate a standardized dataset containing labeled boxes; Step 4: Data augmentation and division of the standardized dataset to construct a training set, a validation set, and a test set; Step 5: Training the YOLO model using the training set, the validation set, and the test set to obtain a trained model; Step 6: Exporting the trained model as an ONNX format model; Step 7: Optimizing the ONNX format model based on TensorRT to generate a TensorRT inference engine file suitable for embedded platforms; Step 8: Deploying the TensorRT inference engine file to an embedded edge computing platform to perform real-time image acquisition, image preprocessing, TensorRT accelerated inference, detection result post-processing, and recognition result output.

2. The method of claim 1, wherein, In step 1, The field environment includes lighting conditions, soil types, and work speeds; The image resolution used during collection ranges from 1280x720 to 1920x1080, the shooting height ranges from 20cm to 80cm, and the shooting frame rate ranges from 15fps to 60fps.

3. The method of claim 1, wherein, In step 2, Data preprocessing and quality screening include: clarity detection of the original image data to eliminate blurred images; exposure detection of the original image data to eliminate severely underexposed or overexposed images; and duplicate and highly similar image deduplication.

4. The method of claim 1, wherein, In step 3, Labeling and quality control include: boundary box labeling and class label assignment for corn seed targets in the qualified image set; multi-person cross-labeling for difficult example samples; and sampling review and consistency verification of the labeling results; Wherein, the minimum labeling target size is 10x10 pixels, and the allowed labeling proportion of occluded targets is greater than or equal to 30% of the visible area.

5. The method of claim 1, wherein, In step 4, Data augmentation of the standardized dataset includes: brightness perturbation, contrast perturbation, color perturbation, random rotation, random scaling, random cropping, and adding noise to simulate field disturbances.

6. The method of claim 1, wherein, In step 5, The input image size of the YOLO model is 1920x1080; the confidence threshold Tc is 0.3-0.7; and the non-maximum suppression threshold Tn is 0.3-0.

6.

7. The method of claim 1, wherein, Step 6 specifically, Perform model structure freezing; export the ONNX intermediate representation; and perform operator compatibility checking and graph structure optimization on the ONNX graph.

8. The method of claim 1, wherein, Step 7 specifically, Select FP16 as the inference precision mode; perform layer fusion and memory optimization; and generate a TensorRT inference engine file suitable for embedded platforms.

9. The method of claim 1, wherein, In step 8, The embedded edge computing platform is the NVIDIA Jetson AGX Orin platform; and the real-time inference speed after deployment is not less than 30 frames per second.

10. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1-9.