An improved YOLOv11 algorithm-based field weed intelligent identification method and device

Abstract

The present application relates to the technical field of agricultural intelligent equipment, and proposes a field weed intelligent identification method and device based on an improved YOLOv11 algorithm, which comprises the following steps: step one, field weed dataset construction and preprocessing; step two, constructing a data configuration file based on the preprocessed dataset, setting YOLOv11 model parameters, and constructing a basic YOLOv11 model; step three, constructing an improved YOLOv11 model; step four, training the improved YOLOv11 model; step five, evaluating the performance of the improved YOLOv11 model; and step six, system deployment and field identification implementation. By introducing an LDConv large-core deep convolution and bidirectional feature pyramid fusion network, and embedding a channel and spatial dual attention mechanism, the present application effectively enhances the model's representation ability for multi-scale weed features and discriminability in complex backgrounds, and improves the detection accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent equipment technology, specifically to a method and device for intelligent identification of field weeds based on an improved YOLOv11 algorithm. Background Technology

[0002] With the rapid development of computer vision and artificial intelligence technologies, deep learning-based visual recognition technology has provided a new technological path for intelligent weed control. Currently, relevant research institutions and enterprises at home and abroad have developed a variety of machine vision-based weed identification systems, but these systems still face many technical bottlenecks and challenges in practical applications.

[0003] The farmland environment is currently highly dynamic and complex, manifested in several aspects: drastic changes in lighting conditions, including strong sunlight on sunny days, diffused light on cloudy days, and backlighting in the morning and evening; crops and weeds exhibit high similarity in color, texture, and morphology, making them particularly difficult to distinguish during the seedling stage; the scale of weed targets varies greatly, from seedlings occupying only a few pixels to mature plants covering hundreds of pixels; mutual occupancy and overlap between plants exist; and the soil background displays different colors and textures under varying humidity and texture conditions. Most existing recognition systems employ general object detection models, such as Faster R-CNN, SSD, and standard versions of the YOLO series. While these models perform well on certain benchmark datasets, they lack specific optimization for the unique farmland scenarios, leading to decreased recognition accuracy and increased false positive and false negative rates in practical applications. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and apparatus for intelligent field weed identification based on an improved YOLOv11 algorithm, which solves the problems of decreased identification accuracy, increased false detection rate, and increased false detection rate of weed identification devices in practical applications.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent weed identification in fields based on an improved YOLOv11 algorithm includes the following steps: Step 1: Construction and preprocessing of the field weed dataset; Step 2: Build a data configuration file based on the preprocessed dataset, set the YOLOv11 model parameters, and build a basic YOLOv11 model; Step 3: Improve the construction of the YOLOv11 model, including the following improvements: Introducing LDConv into the C3k2 module of the Backbone network enhances its multi-scale feature extraction capabilities: In the Neck section, the original PANet is replaced with BiFPN, and the CBAM attention mechanism is embedded to achieve a dual attention mechanism of channel and space. Step 4: Train the improved YOLOv11 model. The training includes: Using the dataset mentioned above, we trained the model in the PyTorch framework and optimized the model parameters using transfer learning and data augmentation strategies. Step 5: Evaluate the performance of the improved YOLOv11 model; Step Six: System Deployment and Field Identification Implementation.

[0006] The preferred field weed dataset includes 12 types of weeds: lettuce, aster, oxalis, goosegrass, sedge, horsetail, purslane, sage, physalis, physalis, peperomia, plantain, ferns and sow thistle, covering different growth stages.

[0007] Preferably, the specific process in step one is as follows: S1: Manual annotation is performed using the LabelImg tool, and the annotation format is a YOLO format txt file; S2: Perform orthorectification, gridding cropping, Gaussian denoising and color enhancement preprocessing on the images, and divide them into training set, validation set and test set in a 7:2:1 ratio.

[0008] Preferably, the performance evaluation of the YOLOv11 model includes: using metrics such as precision, recall, F1 score, and mean precision to evaluate the model's recognition accuracy, robustness, and real-time performance.

[0009] Preferably, the specific process in step six is ​​as follows: S1: Deploy the optimized YOLOv11 model on the Jetson Orin Nano Super development board and use its AI computing power to achieve real-time inference; S2: The development board is integrated with cameras and tracked platforms to build a complete intelligent weed identification system in the field, enabling real-time detection and precise positioning of weeds.

[0010] A field weed intelligent identification device based on an improved YOLOv11 algorithm includes a vehicle body, a wheel support plate fixedly mounted on the vehicle body, two sets of drive wheels rotatably mounted on the wheel support plate, two sets of transmission wheels rotatably mounted on the wheel support plate, tracks wound around the drive wheels and transmission wheels, and two sets of wheel motors fixedly mounted on the wheel support plate, with the output end of each wheel motor fixedly connected to the drive wheel.

[0011] Preferably, a camera body is fixedly installed on the vehicle body, a sensor is fixedly installed on the vehicle body, a display screen is fixedly installed on the vehicle body, and a charging port is provided on the vehicle body.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention effectively enhances the model's ability to represent multi-scale weed features and its ability to distinguish them in complex backgrounds by introducing LDConv large-kernel deep convolution and bidirectional feature pyramid fusion network, and embedding channel and spatial dual attention mechanisms, thereby improving the accuracy and robustness of detection.

[0013] 2. This invention achieves efficient inference and real-time response of the model by relying on the high-performance edge computing power of NVIDIA Jetson Orin Nano Super at the system deployment level, thus meeting the low latency requirements of mobile field operations.

[0014] 3. In terms of environmental adaptability, the integrated tracked mobile platform of this invention has excellent obstacle-crossing and terrain adaptability. Combined with the dustproof and waterproof design of the whole machine, it ensures stable operation and continuous operation in typical farmland environments such as muddy and uneven terrain. Thus, it provides a reliable, efficient and practical technical solution for precision weed control in smart agriculture. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the overall structure of the present invention from a first-view perspective; Figure 2 This is a schematic diagram of the overall structure of the present invention from a second perspective; Figure 3 This is a schematic diagram of a portion of the dataset used in this invention; Figure 4 This is a diagram of the improved YOLOv11 model architecture of the present invention; Figure 5 This is a diagram showing the detection results of the YOLOv11 model for processing field weed datasets according to the present invention; Figure 6 This is a diagram showing the detection results of the YOLOv11 model for processing field weed datasets according to the present invention; Figure 7 This is a diagram showing the detection results of the YOLOv11 model for processing field weed datasets according to the present invention; Figure 8 This is a comparison chart of data before and after the improvement of YOLOv11 in this invention; Figure 9 This is a diagram of the LDConv network structure of the present invention; Figure 10 This is a diagram of the BiFPN feature fusion network of the present invention; Figure 11 This is a diagram illustrating the fusion process of the BiFPN of this invention; Figure 12 This is a structural diagram of channel attention and spatial attention in CBAM of the present invention.

[0016] In the picture: 1. Body of the vehicle; 2. Wheel support plate; 3. Drive wheel; 4. Transmission wheel; 5. Track; 6. Wheel motor; 7. Camera body; 8. Sensor; 9. Display screen; 10. Charging port. Detailed Implementation

[0017] The following will describe in detail the implementation of this application with reference to the accompanying drawings and embodiments, so that the implementation process of how this application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0018] Reference Figures 1-12 This embodiment proposes an intelligent weed identification method based on the improved YOLOv11 algorithm. A field weed dataset is constructed and preprocessed, including 12 types of weeds: Asparagus lettuce, Aster subulatus, Oxalis corniculata, Eleusine indica, Asparagus usurinaria, Goosegrass, Horseweed, Indian aster, Linearleaf Speedwell, Phyllanthus usurinaria, Pilea microphylla, Plantain, Pteris vittata, and Sow thistle, covering different growth stages. The images in the dataset are manually labeled using the image visualization annotation tool LabelImg, with the labels being: Asparagus lettuce, Aster subulatus, Oxalis corniculata, Goosegrass, Horseweed, Indian aster, Linearleaf Speedwell, Phyllanthus usurinaria, Pilea microphylla, Plantain, and Pteris vittata. The labeled information for multifida and sow thistle was saved in txt file to complete the construction of the weed dataset. During the labeling process, the minimum bounding rectangle was used to determine the location of the weeds to ensure the accuracy of the bounding box coordinates and category information. A data configuration file was built based on the preprocessed dataset, and the YOLOv11 model parameters were set to construct a basic YOLOv11 model. Orthorectification, grid cropping, Gaussian smoothing for noise reduction, and color mean square error-based enhancement were performed on the images in the dataset to improve data quality and diversity. Finally, the dataset was randomly shuffled and divided into training, validation, and test sets in a 7:2:1 ratio to construct a representative standardized dataset.

[0019] To address the shortcomings of the original YOLOv11 model in complex field weed detection scenarios, this paper improves the construction of the YOLOv11 model and trains it. The backbone is responsible for extracting rich multi-level features from the input image. To enhance the backbone network's spatial feature perception capability without significantly increasing computational burden, the C3k2 module in the backbone and neck parts introduces the LDConv (Lightweight Deep Convolution) module, forming the C3k2_LDConv module. The LDConv module first performs depthwise convolution (DWConv) on the input feature map to efficiently extract spatial features. Then, a simplified channel attention mechanism is introduced, generating channel weights through global average pooling (GAP) and fully connected layers (FC) to recalibrate the features after deep convolution. This process can be described as follows:

[0020] ·

[0021] Where F is the input feature, Represents depthwise convolution. For global average pooling, It is a fully connected layer. In this way, LDConv can effectively capture local spatial information such as texture and edges of weed leaves with negligible computational cost, and emphasize information-rich feature channels.

[0022] In the Neck section, the original YOLOv11 used PANet for top-down and bottom-up feature fusion. However, in complex weed detection scenarios, weeds vary greatly in scale and often have similar colors and textures to crops and soil backgrounds. PANet's egalitarian feature fusion approach may not adaptively highlight the most critical information for weed detection. Figure 10 As shown, Figure 10 (a) is the original PANet structure diagram. Figure 10(b) is the replaced BiFPN structure diagram. BiFPN introduces a learnable adaptive weight mechanism on the basis of PANet. This mechanism allows the network to dynamically learn and assign importance weights to each input feature when fusing feature maps of different resolutions, rather than simply averaging or splicing them. This enables the network to more selectively strengthen feature layers that contribute more to the current weed detection task, while suppressing redundant or interfering information, thereby achieving more efficient and discriminative feature fusion.

[0023] BiFPN adaptively adjusts the importance of different input features through learnable weights and performs fast bidirectional (top-down and bottom-up) fusion. Assuming the features to be fused are... Then the fusion process of BiFPN can be simplified as follows:

[0024] in , For learnable weights, To prevent small constants from being numerically unstable.

[0025] To further enhance the model's feature selection ability in complex environments, the CBAM attention mechanism is introduced into the Neck part. This mechanism is a lightweight and general attention mechanism that helps the network dynamically adjust its attention to features at different levels during cross-scale information interaction. In particular, it enhances the sensitivity to small-scale or blurry weed features. In complex weed detection scenarios, facing challenges such as similar color and texture between the target and the background, occlusion, and changes in lighting, standard convolutional operations often struggle to autonomously focus on key feature regions. The CBAM module guides the network to adaptively calibrate feature responses through two sequentially connected sub-modules: channel attention and spatial attention. This strengthens features related to weed recognition in the channel dimension and locates the target's location in the spatial dimension, effectively suppressing background interference and improving the model's discrimination ability in complex environments.

[0026] like Figure 12As shown, this module first processes the input features through the channel attention submodule. After global average pooling and global max pooling, the feature maps extract global contextual information and salient features in the channel dimension, respectively. Then, they are fed into a shared multilayer perceptron for fusion and transformation. Finally, a channel weight vector is generated by the Sigmoid function and multiplied with the original input channel by channel to achieve reweighting of the feature channels. The channel-enhanced features enter the spatial attention submodule. This submodule generates two spatial feature maps by performing average pooling and max pooling in the channel dimension, respectively. After concatenating them, a standard convolutional layer is used to model spatial relationships. Then, a spatial weight matrix is ​​generated by the Sigmoid function and finally multiplied with the input features position by position to complete the spatial dimension attention focusing. By introducing LDConv to enhance local feature extraction, using BiFPN to achieve adaptive multi-scale fusion, and embedding CBAM to implement a dual attention mechanism, this embodiment constructs an improved YOLOv11 weed detection model that balances efficiency and accuracy, laying the algorithmic foundation for subsequent real-time field identification tasks.

[0027] To evaluate the performance of the improved YOLOv11 model, weed images collected in the field were acquired. The preprocessed dataset, including those with orthorectification, non-overlapping gridding, Gaussian smoothing for noise reduction, and color mean squared error transformation enhancement, as well as some publicly available datasets, were divided into training, validation, and test sets. These sets were then input into the optimized network model for training. The field weed detection model was validated using metrics such as precision, recall, F1 score, and mean average precision (mAP).

[0028] Precision reflects the proportion of positive samples correctly detected by the model, that is, how many of the detected targets are true targets. The calculation formula is:

[0029] in, This is a true example (the number of weeds that were correctly detected). These are false positives (the number of non-weed targets that are falsely detected). The higher the precision, the lower the false positive rate of the model.

[0030] Recall measures a model's ability to detect real weeds. Recall can be calculated using the following formula:

[0031] Here, FN stands for False Negative; False Negative (FN) refers to the number of real targets that the model failed to detect, representing the number of weeds that the model failed to detect. A recall rate closer to 1 indicates that the model has a higher detection capability when detecting targets, meaning it can better identify real targets. A high recall rate means that the model can capture all real targets well, reducing the possibility of missed detections.

[0032] The F1 score is a metric that comprehensively considers precision and recall, used to evaluate the overall performance of a model. The calculation formula is as follows:

[0033] A higher F1 score indicates a better overall performance of the model in the object detection task.

[0034] Mean precision is a core evaluation metric for object detection. It reflects model performance by calculating the average precision at different recall rates. Common metrics include: mAP50: The mAP value when the IoU threshold is 0.5; mAP50-95: The average mAP value of the IoU threshold from 0.5 to 0.95 (step size 0.05), which has stricter requirements for the accuracy of the detection box.

[0035] The optimized field weed detection model was deployed on the NVIDIA Jetson Orin Nano Super development board, which uses its high-performance AI computing power to achieve real-time weed identification. The development board was installed in a tracked mobile machine with reinforced packaging and connected to a camera to build an intelligent weed identification platform.

[0036] A smart weed identification device for fields based on an improved YOLOv11 algorithm includes a vehicle body 1, a wheel support plate 2 fixedly mounted on the vehicle body 1, two sets of drive wheels 3 rotatably mounted on the wheel support plate 2, and two sets of transmission wheels 4 rotatably mounted on the wheel support plate 2. Tracks 5 are wound around the drive wheels 3 and transmission wheels 4, connecting them via the tracks 5 to ensure good passability and stability in complex field terrain. Two sets of wheel motors 6 are fixedly mounted on the wheel support plate 2, with the output of each wheel motor 6 fixedly connected to the drive wheels 3 to provide power for the tracks 5. A camera body 7 is fixedly mounted on the vehicle body 1 for collecting field image information. Sensors 8 are fixedly mounted on the vehicle body 1 for environmental perception. A display screen 9 is fixedly mounted on the vehicle body 1 for displaying working status and identification results. A charging port 10 is provided on the vehicle body 1 for powering the device. The device uses NVIDIA Jetson Orin... NanoSuper serves as the core processing unit, equipped with the improved YOLOv11 weed recognition algorithm, enabling real-time and accurate detection and location of weeds in the field. The entire machine is dustproof, waterproof, autonomously navigates, and can operate continuously, making it suitable for intelligent plant protection and field management operations in precision agriculture.

[0037] Working principle: The device uses NVIDIA Jetson Orin Nano Super as the core processing unit and is equipped with an improved YOLOv11 weed recognition algorithm, which can realize real-time and accurate detection and positioning of weeds in the field. The whole machine is dustproof and waterproof, autonomous navigation and continuous operation capabilities, and is suitable for intelligent plant protection and field management operations in precision agriculture.

[0038] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification of field weeds based on an improved YOLOv11 algorithm, characterized in that, Includes the following steps: Step 1: Construction and preprocessing of the field weed dataset; Step 2: Build a data configuration file based on the preprocessed dataset, set the YOLOv11 model parameters, and build a basic YOLOv11 model; Step 3: Improve the construction of the YOLOv11 model, including the following improvements: Introducing LDConv into the C3k2 module of the Backbone network enhances its multi-scale feature extraction capabilities: In the Neck section, the original PANet is replaced with BiFPN, and the CBAM attention mechanism is embedded to achieve a dual attention mechanism of channel and space. Step 4: Train the improved YOLOv11 model. The training includes: Using the dataset mentioned above, we trained the model in the PyTorch framework and optimized the model parameters using transfer learning and data augmentation strategies. Step 5: Evaluate the performance of the improved YOLOv11 model; Step Six: System Deployment and Field Identification Implementation.

2. The intelligent field weed identification method based on the improved YOLOv11 algorithm according to claim 1, characterized in that, The field weed dataset includes 12 categories of weeds, covering different growth stages: lettuce, aster, oxalis, goosegrass, sedge, horsetail, sage, sage, physalis, peperomia, peperomia var. chinensis, plantain, fern, and sow thistle.

3. The intelligent field weed identification method based on the improved YOLOv11 algorithm according to claim 1, characterized in that, The specific process in step one is as follows: S1: Manual annotation is performed using the LabelImg tool, and the annotation format is a YOLO format txt file; S2: Perform orthorectification, gridding cropping, Gaussian denoising and color enhancement preprocessing on the images, and divide them into training set, validation set and test set in a 7:2:1 ratio.

4. The intelligent field weed identification method based on the improved YOLOv11 algorithm according to claim 1, characterized in that, The performance evaluation of the YOLOv11 model includes using metrics such as precision, recall, F1 score, and mean precision to assess the model's recognition accuracy, robustness, and real-time performance.

5. The intelligent field weed identification method based on the improved YOLOv11 algorithm according to claim 1, characterized in that, The specific process in step six is ​​as follows: S1: Deploy the optimized YOLOv11 model on the Jetson Orin Nano Super development board and use its AI computing power to achieve real-time inference; S2: The development board is integrated with cameras and tracked platforms to build a complete intelligent weed identification system in the field, enabling real-time detection and precise positioning of weeds.

6. A field weed intelligent identification device based on an improved YOLOv11 algorithm, comprising a vehicle body (1), characterized in that, A wheel support plate (2) is fixedly installed on the main body (1). Two sets of power wheels (3) are rotatably installed on the wheel support plate (2). Two sets of transmission wheels (4) are rotatably installed on the wheel support plate (2). Tracks (5) are wound around the power wheels (3) and transmission wheels (4). Two sets of wheel motors (6) are fixedly installed on the wheel support plate (2), and the output end of each wheel motor (6) is fixedly connected to the power wheel (3).

7. A field weed intelligent identification device based on an improved YOLOv11 algorithm according to claim 6, characterized in that, A camera body (7) is fixedly installed on the vehicle body (1), a sensor (8) is fixedly installed on the vehicle body (1), a display screen (9) is fixedly installed on the vehicle body (1), and a charging port (10) is provided on the vehicle body (1).

Images