CN-121999203-A - Park weed accurate identification method and system based on improved YOLOv-11 deep convolutional network

CN121999203ACN 121999203 ACN121999203 ACN 121999203ACN-121999203-A

Abstract

The invention discloses a park weed accurate identification method and a park weed accurate identification system based on an improved YOLOv11 deep convolution network, comprising the following steps of firstly, constructing a multi-scene park weed image data set, secondly, carrying out data enhancement and pretreatment, thirdly, constructing an improved YOLOv11 lightweight network model, fourthly, designing a self-adaptive loss function, fifthly, executing multi-stage model training and optimization, and sixthly, realizing real-time weed identification and result output. According to the invention, by introducing a light-weight attention mechanism and an improved multi-scale feature fusion structure, the capability of extracting and distinguishing small-scale weed features under a complex background by a network is remarkably enhanced, the false detection rate and the omission factor are effectively reduced, and the adopted dynamic label distribution strategy and the loss function fused with global context information promote the detection robustness and the positioning accuracy of the model under dense, shielding and morphological changeable weed scenes.

Inventors

BI ZHIQIANG
ZHANG YUCHENG
Yang Zhicun
Xie Gangzhu
ZHANG MENG
XIONG SHILIN
WANG DONGZHAO
HU LUJUN

Assignees

北京国科廪科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (9)

1. The park weed accurate identification method based on the improved YOLOv-depth convolution network is characterized by comprising the following steps of: Firstly, constructing a multi-scene park weed image data set, namely acquiring RGB images of park greening areas under different illumination, seasons and weather conditions through image acquisition equipment carried on a mobile platform, and marking various weed targets in the images by using a boundary frame to form an initial data set; step two, data enhancement and preprocessing, namely, performing enhancement operations comprising random rotation, brightness contrast adjustment, simulation shielding and background mixing on the images in the initial data set, uniformly scaling all the images to a fixed size, and performing normalization processing to generate an enhancement training set; Step three, an improved YOLOv light-weight network model is constructed, namely a light-weight channel attention module is embedded behind a key feature extraction layer of a backbone network of the light-weight network model by taking the YOLOv network as a reference, meanwhile, the structure of a neck feature pyramid network is adjusted, a branch from a shallow feature map is added, and the feature expression capability of small-scale weeds is enhanced by a weighted bidirectional feature fusion mode; Step four, designing a self-adaptive loss function, namely constructing a composite loss function integrating the regression loss of the bounding box, the confidence loss of the category and the focusing loss of the difficult sample, and introducing a weighting factor considering a target scale into the regression loss of the bounding box so as to improve the positioning precision of the model on the dense weeds; Training the improved YOLOv lightweight network model by using the enhanced training set, wherein a strategy of freezing a backbone network is adopted in the early training stage, and end-to-end adjustment is performed by thawing in the later training stage; Inputting park real-time video stream or image to be detected into the weed detection model, outputting recognition result containing weed category, confidence and accurate boundary frame coordinates in the image by the model, and transmitting the recognition result to a subsequent decision system.
2. The park weed accurate identification method based on the improved YOLOv-depth convolutional network according to claim 1, wherein the lightweight channel attention module in the third step adopts global average pooling to generate channel weights, calculates attention vectors through a full-connection layer structure comprising dimension reduction and dimension elevation, and finally performs channel-by-channel multiplication weighting with the original feature map without introducing additional pooling branches to keep module lightweight.
3. The park weed accurate identification method based on the improved YOLOv-11 deep convolutional network, which is disclosed by claim 2, is characterized in that the weighted bidirectional feature fusion mode in the step three is specifically that a learnable adaptive weight parameter is introduced for input feature graphs from different levels in an up-sampling fusion path and a down-sampling fusion path of a feature pyramid, and the contribution degree of each feature graph is dynamically optimized in the network training process.
4. The method for accurately identifying park weeds based on the improved YOLOv-depth convolutional network according to claim 1, wherein in the fourth step, the calculation mode of the weighting factor considering the target scale is associated with the area ratio of the prediction boundary box and the real boundary box, so that the model is given higher regression loss weight to the weed targets with smaller area in the training process.
5. The method for accurately identifying weeds in parks based on the improved YOLOv-deep convolutional network according to claim 4, wherein the multi-stage model training in the fifth step further comprises a knowledge distillation step, namely a teacher model which is pre-trained on a general data set and has higher precision is used for guiding a student model with lower precision, namely the improved YOLOv-lightweight network model, and the identification precision and generalization capability of the student model are improved by minimizing distillation loss between the teacher model and an output feature map of the student model.
6. The method of claim 5, wherein the selected teacher model is a YOLOv-X or larger visual transducer model, wherein distilling losses include feature modeling losses based on the teacher and student model intermediate layer feature maps, and wherein predicting KL divergence losses for the output soft labels based on both.
7. The method for accurately identifying park weeds based on the improved YOLOv-depth convolutional network according to claim 1, further comprising the steps of seventh, performing non-maximum suppression screening on an original detection frame output by a model, and performing post-processing and geographic mapping, wherein the screened image coordinates are converted to actual positions under a park geographic coordinate system through camera parameters calibrated in advance, so that a weed distribution map with geographic information is generated.
8. Park weed accurate identification system based on improvement YOLOv11 degree of depth convolutional network, characterized by comprising the following modules: The image acquisition module is used for acquiring a real-time video stream or an image sequence of the park greening area through camera equipment arranged on the mobile platform or the fixed monitoring point; The data preprocessing module is in communication connection with the image acquisition module and is used for performing size scaling, normalization and format conversion on the input image so as to enable the input image to meet the model input requirement; Improving YOLOv a weed detection model module, loading an optimal weed detection model obtained through training, and receiving the output of the data preprocessing module, and performing forward reasoning to generate a preliminary weed identification result; The post-processing and outputting module is in communication connection with the improved YOLOv weed detection model module and is used for performing confidence degree filtering and non-maximum value suppression operation on the primary identification result and formatting and outputting final weed category, confidence degree and position information; And the system control and integration interface module is in communication connection with the post-processing and output module and is used for coordinating the working flow of each module and sending the final identification result to a display terminal or a control system of the automatic weeding equipment through a communication interface.
9. The park weed precision identification system based on the improved YOLOv-depth convolutional network of claim 8, further comprising a geographic information mapping module receiving the output of the post-processing and output module for converting weed position information under an image coordinate system to a park geographic information system coordinate system in combination with real-time positioning information and camera calibration parameters of the mobile platform.

Description

Park weed accurate identification method and system based on improved YOLOv-11 deep convolutional network Technical Field The invention relates to the technical field of computer systems, in particular to a park weed accurate identification method and system based on an improved YOLOv-depth convolution network. Background Currently, greening maintenance management of urban parks is developing to be fine and intelligent. The accurate identification and positioning of weeds are the precondition and key of automatic weeding operation. The traditional method is mostly dependent on manual inspection or recognition technology based on simple image processing, has low efficiency and is easily interfered by complex background, illumination change and morphological similarity of weeds and ornamental plants. In recent years, a target detection algorithm based on deep learning, particularly a YOLO series, has great potential in the field of agricultural plant protection. However, when the general YOLO model is directly applied to park weed identification scenes, a plurality of challenges are faced, namely, park environment backgrounds are complex and changeable, weed targets are usually small and densely distributed, and color and texture confusion exists between the park environment backgrounds and the grass or flowers. The existing model has the problems of high omission rate for small target weeds, high false detection rate for similar plants, large quantity of model parameters, being unfavorable for deployment at the mobile weeding equipment end and the like. Therefore, a light-weight and high-precision weed identification method and system for optimizing and improving a park complex scene are needed to meet the requirements of precise and efficient intelligent garden management. Disclosure of Invention Therefore, the invention provides a park weed accurate identification method and system based on an improved YOLOv-depth convolution network, which are used for solving the problems that in the prior art, the small target weed detection omission rate is high, the similar plant false detection rate is high, and the model parameters are large and are unfavorable for being deployed at a mobile weeding equipment end. In order to achieve the above object, the present invention provides the following technical solutions: Park weed accurate identification method based on improved YOLOv-depth convolution network comprises the following steps: Firstly, constructing a multi-scene park weed image data set, namely acquiring RGB images of park greening areas under different illumination, seasons and weather conditions through image acquisition equipment carried on a mobile platform, and marking various weed targets in the images by using a boundary frame to form an initial data set; step two, data enhancement and preprocessing, namely, performing enhancement operations comprising random rotation, brightness contrast adjustment, simulation shielding and background mixing on the images in the initial data set, uniformly scaling all the images to a fixed size, and performing normalization processing to generate an enhancement training set; Step three, an improved YOLOv light-weight network model is constructed, namely a light-weight channel attention module is embedded behind a key feature extraction layer of a backbone network of the light-weight network model by taking the YOLOv network as a reference, meanwhile, the structure of a neck feature pyramid network is adjusted, a branch from a shallow feature map is added, and the feature expression capability of small-scale weeds is enhanced by a weighted bidirectional feature fusion mode; Step four, designing a self-adaptive loss function, namely constructing a composite loss function integrating the regression loss of the bounding box, the confidence loss of the category and the focusing loss of the difficult sample, and introducing a weighting factor considering a target scale into the regression loss of the bounding box so as to improve the positioning precision of the model on the dense weeds; Training the improved YOLOv lightweight network model by using the enhanced training set, wherein a strategy of freezing a backbone network is adopted in the early training stage, and end-to-end adjustment is performed by thawing in the later training stage; Inputting park real-time video stream or image to be detected into the weed detection model, outputting recognition result containing weed category, confidence and accurate boundary frame coordinates in the image by the model, and transmitting the recognition result to a subsequent decision system. Preferably, the light-weight channel attention module in the third step adopts global average pooling to generate channel weight, calculates the attention vector through a full-connection layer structure comprising dimension reduction and dimension increase, and finally performs channel-by-channel multiplication weighting with the original