Research On Automatic Counting Of Wheat Grain Based On Image Processing

Seeds are like the "chips" of agricultural cultivation.They are the material foundation for the development of a country.They can reflect the real strength and influence of a country and also reflect the responsibility of a country.In order to improve crop yield and improve crop varieties,it is necessary to evaluate the quality of seeds.The commonly used index is the thousand seed weight of seeds.Counting is an important step to measure the thousand seed weight of seeds.Manual counting and photocell counting were common seed counting methods in the early stage,which had many problems such as complicated operation,high cost and low efficiency.The automatic crop counting method studied in this paper takes wheat seeds as the experimental research object and studies the automatic counting method of wheat grains based on machine vision image processing technology.The details are as follows.Firstly,the automatic counting of wheat seeds based on traditional image segmentation morphological algorithm and improved watershed algorithm was implemented.The specific implementation is as follows: the wheat image is preprocessed,including basic operations such as binarization,grayscale and filling,then the image processing method is segmented,and finally the connected area is extracted to realize counting.For the morphological segmentation algorithm,the counting accuracy of slightly adherent seeds could be improved by changing the corrosion intensity.For watershed segmentation algorithm,the counting accuracy of watershed algorithm is improved by adding distance transformation and expansion corrosion operation.These two methods are improved on the basis of the original method,and improve the counting accuracy of wheat seeds.For the image of wheat seeds without adhesion,it can basically count accurately,but for some deeply adhesion and stacked seeds,it can not be completely segmented and can not meet the actual demand of seed counting.Secondly,the automatic counting of wheat grains based on deep learning YOLOv4 algorithm was realized,and a large number of wheat seed images were collected to make data sets.TTlabel software was used to label the wheat seeds in the image,and then the established YOLOv4 model was used to train the data set.Finally,three algorithms are used to recognize and count images with different lighting,seed density,adhesion degree and background,and the accuracy of recognition and counting of the three algorithms is compared and analyzed.The comparison results show that the accuracy of counting wheat seed images with different light conditions,non-adhesion,low density and single color background can reach 85% by using morphology and improved watershed algorithm.However,the seed images with conglutination,stacking and dark mixed background cannot be completely segmented and the counting error is large.When YOLOv4 algorithm based on deep learning was used to process wheat seed images under different conditions such as adhesion and stacking,the accuracy of recognition and counting was 98.41%,the recovery rate was 98.66%,the accuracy was 99.42%,the F1 value was99.04% and the average accuracy was 98.41%.It solves the problem of seed counting under different conditions such as severe adhesion and stacking,and can meet the actual needs of seed counting with strong transferability.