| Wheat is one of the three major food crops and plays a very important role in ensuringnational food security.Its spike phenotypic parameters also provide decision support for accurate breeding and intelligent cultivation of wheat.The panicle seed number can reflect the quality of wheat varieties and can estimate the yield.However,the traditional panicle seed of wheat count is mainly conducted by threshing,which destroys the integrity of the wheat spike and is not conducive to the study of phenotypic traits of wheat spike;the spike length and spike width are important characteristics of wheat spike phenotypes,but the traditional measurement methods are mostly manual,which is time-consuming and laborious.Therefore,in this thesis,based on the image characteristics of the wheat spike,deep learning and image processing methods were used to obtain the number of seeds in the wheat spike and the spike length and spike width of wheat,respectively.The specific work and research findings are as follows:(1)To construct a wheat spike dataset and propose a method for counting wheat seeds on the spike.Five varieties of wheat were collected and de-stalked,and 1500 images of wheat spikes were taken with a mobile phone.After stripping and analysing the wheat samples and combining the images of the wheat spikelets,it was found that the number of transparent region in the wheat spikelet can reflect the number of wheat spikelet grains to a certain extent,and based on this rule and the actual situation of the wheat varieties,a wheat in-spike seed counting method was proposed:The spikelets of wheat are divided into four categories,i.e.1 to 4 seeds.For a single wheat spikelet,the pictures of the front and back of the spikelet are taken,and the number of seeds in each of the four categories is found on both sides and multiplied and added to the number of seeds represented by the corresponding category to obtain the number of seeds in a single wheat spikelet,and about11,700 spikelets were labeled according to this method.(2)Building and improving the panicle seed of wheat recognition counting model.Four target detection network models of YOLO series were constructed and trained and compared respectively.The YOLOv7 model was then improved by fusing the attention mechanism with the YOLOv7 model,which has a relatively good overall performance.The experimental results show that the optimized YOLOv7-CBAM model has the best performance with a mean average precision of 66.51%,which is 2.92%higher than the original YOLOv7 model.The model was applied to the panicle seed of wheat recognition counting,and compared with the actual number of seeds of five wheat ears measured manually,the average absolute error was 4.82 and the average relative error was 14.42%,which provided a reference for wheat seed counting.(3)Measurement of wheat spike length and spike width.A method was introduced to measure the geometric parameters of the wheat spike by means of the scale factor of the reference,using Open CV to denoise the image of the wheat spike,then using image morphological operations to remove the awn and stalk of the wheat spike,and finally measuring the spike length and width of the wheat spike by finding the minimum outer rectangle.By comparing the manual statistics,the average relative errors of the measurements of wheat spike length and spike width were 1.62%and 3.58%respectively,and the R~2 of the exponential regression model was 0.982 and 0.831 respectively,which verified the accuracy of this method.(4)Design and implementation of a graphical interface for the acquisition of phenotypic parameters in wheat spikes.Design of a graphical interface for the acquisition of phenotypic parameters of wheat spikes using the Py Qt5 GUI tool Qt Designer and the Python language,enabling the panicle seed of wheat of single wheat plants of to be counted as well as batch counting and file saving;The functions for measuring the spike length and width of wheat are also integrated into the graphical interface. |
PDF Full Text Request