Wheat Ear Recognition And System Implementation Based On Image Processing And Deep Learning

Spike number per unit area is an important factor affecting wheat yield.It is important to determine spike number quickly for estimating wheat yield.Manual counting is often used in production to estimate the output,which is time-consuming and laborious.Therefore,it is of great significance to develop a fast and accurate method to obtain spike number for wheat yield estimation.In this paper,UAV and mobile devices are used to obtain the field ear image by vertical shooting.The ear image is preprocessed by image enhancement and denoising.K-means clustering is used to realize the automatic accurate positioning and segmentation of the ear.After the segmentation,the ear is sorted,and four kinds of tags,including non ear,one ear,two ears and three ears,are extracted to construct the data set After data enhancement and data standardization,the constructed data set is sent to the constructed convolutional neural network(CNN)model for training and verification.The trained model is used to recognize and count wheat ears.Finally,the wheat ear recognition process is integrated into the web system.The results are as follows:1.Construction and verification of CNN model.After the segmentation,more than 140000 wheat images were generated,and four kinds of tags,including non wheat,one wheat,two wheat and three wheat,were selected to construct data sets,which were 8000,8000,2173 and 1193 respectively.The four kinds of data sets were expanded to 12000 by means of data enhancement such as flipping and rotation,in which the training set was 11000 and the verification set was 1000.The experimental results show that the recognition accuracy of non wheat ear,one wheat ear,two wheat ears and three wheat ears is 99.8%,97.5%,98.07% and 98.5% respectively.2.Compare and analyze the recognition results of wheat images taken by different equipment.In this paper,60 mobile phone images and 20 UAV images are selected for recognition.Among them,the recognition result R2 of mobile phone is0.97,RMSE is 0.03,and the recognition result R2 of UAV is 0.97,RMSE is 9.47.The results show that the two methods can effectively identify the ears of wheat,and the photos taken by mobile phone are slightly better than those taken by UAV.3.The recognition results of wheat ear images taken from May 6 to 20 are compared and analyzed.By comparing the wheat ear recognition results of different dates in Xuchang,it can be seen that the performance of wheat ear recognition is different at different stages of wheat filling stage.The root mean square error(RMSE)between the wheat ear recognition results of CNN model and the artificial wheat ear count decreases from 22.54 to 3.24.Among them,the ear images collected in the late filling stage of wheat ear on May 20 achieved the best recognition effect(R2 = 0.99,RMSE = 3.24),which is in good agreement with the early research.4.Select different varieties of wheat for comparative analysis of identification results.In view of the difference of ear characteristics of different varieties of wheat may affect the recognition effect of ear,this paper selected 50 photos of ear,a total of10 ear varieties from the experimental area of Yuanyang farm of Henan Agricultural University to count ear recognition,in order to verify whether the model has good recognition effect among different wheat varieties.The results showed that the accuracy of 50 different varieties of wheat ear recognition ranged from 92.70% to99.63%,the high consistency among 10 varieties was observed,and the standard error was about 12.43.The model had a certain stability in the application of different varieties of wheat ear recognition and counting.5.The system uses Java to realize web function,python thrift server to integrate CNN model,opencv and other technologies to realize ear recognition function,and uses thrift framework to realize RPC interaction to complete the main functions.The system can not only recognize and count the wheat images,but also provide the recognition and error correction function of wheat images and the relearning function of CNN model.