Counting Algorithm Of Soybean Seedlings Based On Object Detection And Density Estimation

Soybean is a kind of globally important food crop,and increasing soybean yields is an urgent need for agriculture in China today.Early field management is indispensable for increasing soybean production.Fast and accurate seedling emergence rate is helpful for farmers to inter-plant and replenish seedlings in a timely manner,as well as to evaluate the quality of seeds.At present,the statistical work of soybean seedling emergence rate is mainly done by manual observation and recording,which takes a lot of time and wastes human and material resources.Therefore,under the general trend of combining agriculture and computer technology,the statistical process of soybean seedling emergence rate needs to be further improved in terms of information and intelligence.In order to quickly obtain soybean seedling emergence rate,timely interplanting,seedling replenishment and evaluation of seed quality,and thus improve soybean yield.In this paper,we propose a method for counting soybean seedlings based on target detection and density estimation methods,with the aim of obtaining information on soybean seedling emergence rate quickly and accurately.The main work of this paper is as follows:(1)Production of soybean seedling dataset.In this paper,a DJI M600 pro UAV was used to take aerial photos of soybean seedlings,and the obtained image data were filtered,spliced,and cropped to obtain usable images of soybean seedlings.The seedlings in the images were manually counted and labeled to obtain the soybean seedling dataset.Traditional data enhancement methods were used to expand the dataset,and data enhancement methods of histogram equalization and brightness enhancement were used.It was experimentally demonstrated that all three methods were effective in improving the recognition performance of the model,and the most performance improvement was achieved when all of them were used.(2)A YOLO v5s-based algorithm for counting soybean seedlings was proposed.First,YOLO v5 s was selected as the base model,Bi FPN and CBAM modules were introduced,and the detection head of the original algorithm was replaced with Transformer Predict Head for improving the detection performance of the model.Experiments are conducted on soybean seedling dataset to compare other target detection methods,the MAE is 1.73 and RMSE is 4.57.The results demonstrate the effectiveness and superiority of the algorithm.(3)A density estimation soybean seedling counting algorithm incorporating convolutional attention is proposed.Firstly,the counting problem is regarded as a mapping relationship with the density map.In order to generate high-quality density maps,the front-end network of the original model is replaced with vgg19 on the basis of CSRNet to improve the network depth and feature extraction capability.The convolutional attention module is also introduced to further improve the counting performance of the model.Finally,pixel point probability summation is performed on the generated density map,and the predicted number of soybean seedlings can be obtained.Experiments were conducted on the soybean seedling dataset to compare other density estimation methods,the MAE is 2.35 and RMSE is 4.69.The results demonstrated the effectiveness and superiority of the algorithm.(4)A soybean seedling counting system was designed and implemented.The system is deployed on a PC and integrates the two soybean seedling counting algorithms proposed in this paper.Through a simple interface operation,the user can select the model and soybean seedling images for soybean seedling count or add a new dataset for training a new model,and the data obtained from the count can be stored in a database.The feasibility of the system has been experimentally verified.This paper combines object detection methods and density estimation methods in computer vision to achieve fast counting of soybean seedlings.A new method is provided to count the seedling emergence rate of soybean,which reduces human labor intensity and saves labor cost.Also,it provides some ideas for the seedling emergence rate acquisition needs of other crops.