| The poor emergence,uneven growth,and weed stress in the field will hinder crop growth.Rapid detection of soybean seedling status and weed distribution,and accurate management of soybean seedlings in the field are important to ensure soybean production.Based on the low-altitude UAV platform,this paper took soybean seedlings(VC stage to V2 stage)as the object,and combined machine vision and deep learning technology to detect the distribution information of crops and crop rows in the field and realized the status of soybean seedlings and weeds distribution information in the field.At the same time,based on the key detection algorithm which were proposed in this paper,a soybean seedling situation detection system is developed,which realizes the rapid detection and analysis of soybean seedling situation based on UAV images.(1)Soybean leaf recognition based on deep learning target detection technology.In this paper,by comparing the accuracy and speed of different models,YOLOv5 s model had been chosen to identify soybean leaf targets.Combined with the characteristics of the UAV field images and the structure of the model itself,this paper simplifies and optimizes the YOLOv5 s model,and compared the model before and after optimization.The model test results shows that the optimized YOLOv5 s model has an accuracy of 91.8%,and mAP@0.5 reaches 0.938.Compared with the original YOLOv5 s model,mAP@0.5 only reduced by 0.029.The average single recognition speed of the optimized model is only 6.42 ms,155.76 frames per second,which is 40% faster than the original model.The optimized YOLOv5 s model in this paper has realized performance improvement with the premise of maintaining the accuracy.(2)A crop line detection method for field UAV images is proposed.By analyzing the geometric arrangement characteristics of crop lines in the image,crop line detection is divided into four processes: geometric transformation,filtering,segmentation and line fitting.Geometric transformation regards the crop row as a linear texture,detects the crop row texture direction in the image through Radon transformation,and constrains the image crop row texture direction to the crop row along the X-axis distribution,on this basis,the long axis is parallel to the crop row direction.The long grid filters noise outside the crop rows.At the same time,in order to realize automatic crop row segmentation,based on the traditional k-means clustering,a method for automatically determining the K value is designed according to the arrangement law of crop rows.Finally,the least squares line fitting method is used to fit the segmented crop rows respectively.which realizes the rapid detection of crop row distribution in UAV images.In the randomly selected verification data,the recognition rate of the centerline of the crop row reaches 93.9%,and the recognition rate reaches more than 92% under different planting modes and shooting heights of images.The average angle deviation compared with manual visual inspection is 0.126°,and the detection time is 0.31 seconds.The method has good performance in both accuracy and speed.(3)The extraction method of key indicators of soybean seedling status was designed.After comparing multiple indices,the Ex G-Ex R was used to binarize the distribution of ground vegetation,and combined with the identification results of YOLOv5 s,soybean leaves were also segmented.Based on the distribution of ground vegetation,soybean leaves and crop rows,calculation methods for key indicators of seedling condition detection is designed: a)Based on the soybean leaf distribution and crop row distribution images,small-area operations are performed on the crop distribution on the crop row to obtain the coefficient of variation of ground crop canopy coverage in different regions was obtained,and the quantitative assessment of soybean seedling distribution was realized.b)Based on the binarized image of a single crop row,the projection segmentation method is used to locate the position of missing seedlings in the crop row,which realizes the detection of the missing seedling position and the length of the broken row in the image,and the calculation of the missing seedling rate and the emergence rate.Planting agronomic and image shooting parameters were combined to estimate the number of normal and missing seedlings in the image.c)By subtracting the distribution of soybean leaves from the distribution of the ground vegetation canopy,combined with open operation filtering and small area operation,the UAV image was detected,and the field weed distribution image was drawn.(4)Designed and developed a field monitoring system for seedling condition.Based on Qt and Python,a field monitoring system was designed and developed that integrates field data display,processing,analysis and other functions.Focusing on the main detection algorithm requirements,the system designed data management module,core detection module,data post-processing module and other modules,which realize the fast whole-process operation from the loading of the original data to the output of the final result. |
