| Rapeseed is a crop that requires a large amount of fertilizer and has strong nutrient tolerance.As one of the main organs of crop,the growth status of leaves can reflect the nutritional status of the crop.By monitoring the growth status of leaves under different stresses,the plant growth pattern can be better understood,and resources can be used more reasonably.However,existing detection methods such as manual measurement and tensile measurement have problems such as high labor intensity,time-consuming,difficult to sustain,and low accuracy.In this paper,based on the growth characteristics of rapeseed leaves,a non-destructive rapeseed leaf image acquisition device was designed and built using machine vision and deep learning technology.A leaf morphological feature extraction method was proposed,and a missing element rapeseed diagnosis model was constructed to achieve accurate monitoring of rapeseed leaf growth and nutritional status diagnosis.The main research content and conclusions are as follows:(1)Design and experiment of rapeseed leaf device.To non-destructive acquisition of rapeseed leaf image information,a rapeseed leaf non-destructive acquisition device based on machine vision technology was designed.The device consists of a laptop,an industrial camera,and a lighting control system,and the industrial camera and lighting system are controlled by sending commands from the computer to achieve nondestructive acquisition of leaf images.The device-acquired images were compared with traditional manual measurement and existing tensile measurement methods,using the area obtained from a scanner and professional leaf processing software as the standard value.The results showed that the detection errors of traditional manual measurement and tensile measurement methods were 9.13% and 9.88%,respectively,the error of the clamp measurement method designed in this paper was 3.60%.The rapeseed growth monitored by the clamp measurement device designed in this paper was close to the natural growth state of rapeseed,indicating that the designed device realized the nondestructive acquisition of rapeseed leaves,providing technical support for the collection of data sets in subsequent segmentation model training.(2)Proposed rapeseed leaf segmentation model based on semantic segmentation.A total of 1080 rapeseed images were collected using a continuous image acquisition device for rapeseed leaf image information.To ensure the robustness of the deep network model,the rapeseed dataset was enhanced with operations such as cropping,mirroring,stretching,and rotating,resulting in 18,360 enhanced images.Several commonly used semantic segmentation network models were compared on the oilseed image dataset,and the UNet network model was selected as the segmentation network backbone due to its good performance.The SENet attention mechanism was added to the network to enhance the ability to extract fine-grained features.The feature layer with added attention mechanism was visualized,and the results showed that the attention mechanism effectively reduced the recognition of useless features and increased the recognition of leaf features.The impact of CE,Focal,and Dice loss functions and their combinations as loss functions for oilseed leaf segmentation tasks was compared.It was demonstrated that the CE+Dice loss function combination could enable the model to converge faster and reduce the time required for training the model.An algorithm for extracting leaf morphological parameters was written in Python language,and the algorithm was used to extract the leaf length,width,area,and perimeter from the segmented images.(3)Proposed oilseed deficiency diagnosis model based on deep learning.A deficiency diagnosis model for oilseed leaves was proposed based on deep learning,which combines multi-scale feature extraction and attention mechanism to address the difficulty of manually identifying oilseed leaf deficiency symptoms.A feature extraction module consisting of multiple scale convolution kernels was proposed to improve the feature perception effect of Res Net50 network on oilseed leaves,and the multi-scale feature fusion optimized the feature transfer between residual convolution structures.The attention mechanism was also added to concentrate the model’s feature extraction on the leaf area,highlight the information features selectively,and achieve precise feature positioning.The results showed that the multi-scale feature fusion module and attention mechanism effectively improved the extraction of image feature information.Compared with VGG and Mobile Net V2 networks,the CBAM-MutiRes Net50 network improved accuracy by 1.1% and 11.3%,precision by 2.2% and 18%,recall by 1.6% and 17.5%,and F1-Score by 2.1% and 18.3%,respectively.For oilseed leaves with deficiency symptoms,the recognition accuracy reached 99.4%,and the model could effectively identify the differences between deficiency symptoms,meeting the high-precision identification requirements of oilseed leaf deficiency images and providing reliable evidence for crop leaf deficiency discrimination with wide application range.(4)Design of rapeseed leaf detection software.Using Py QT5 to create the software interface,integrating leaf morphology parameter extraction model and rapeseed deficiency diagnosis model.The leaf morphology parameter extraction model and rapeseed deficiency diagnosis model were integrated into the software interface using the Python programming language on a PC.The software is capable of extracting leaf parameters and diagnosing rapeseed leaf deficiencies.Users simply need to select an image on their computer and click the corresponding button to obtain results.The software interface is simple,the operation is easy,and the results are accurate,providing users with a fast and non-destructive detection tool for rapeseed leaves.This article designs a rapeseed leaf monitoring system,which includes a rapeseed leaf image collection device used to collect rapeseed images.The system combines a leaf semantic segmentation model and establishes a rapeseed leaf morphology feature extraction model and a rapeseed deficiency diagnosis model.This study is a new exploration of non-destructive monitoring technology for rapeseed growth status,providing new ideas and technical support for intelligent monitoring of rapeseed. |
