Research On Seed Variety Identification Based On Deep Learning And Hyperspectral Imaging Technology

Seed variety and quality are key factors determining the yield of agricultural planting,which in turn affects the rural economic development and national food security.Seed variety identification is the key to ensure seed quality and seed resource management,and plays an important role in effectively controlling purity of the seed,quality and yield of crop products,and ensuring seed breeding safety and seed industrialization.To circumvent the problems in seed variety identification,hyperspectral imaging,image processing and machine learning are combined in this thesis to construct identification model of melon seed varieties and identification method of multi-category crop seeds.The main research contents and conclusions are as follows:(1)Constructing identification model of melon seed hyperspectral data based on machine learning.First,a visible light hyperspectral camera is used to collect melon seed hyperspectral images,the snake image segmentation algorithm is used to extract the contour information of melon seeds,and then the spectral reflectance of seeds is calculated.Based on the hyperspectral data of melon seeds,seed variety discrimination models based on support vector machine,linear discriminant analysis and convolutional neural networks are constructed.In addition,this thesis investigates the influence of preprocessing methods such as SG smoothing algorithm,SG smoothing combined with first derivative,multiple scattering correction and the characteristic wavelength selection methods such as continuous projection algorithm and principal component analysison on the model accuracy.It is found that discrimination models based on the full-band spectrum have the highest accuracy.For the linear discriminant analysis model established based on the full wavelength,for example,the discrimination accuracy rate of both the training set and the prediction set reaches 100%.As the continuous projection algorithm and principal component analysison are used to select characteristic wavelength,the model accuracy is reduced.Combined with the multiple scattering correction data preprocessing and principal component analysis algorithms,the accuracy of the training set and test set of the support vector machine discriminant model still reach at 99.88% and 99.23%respectively when the number of features is reduced by 56.7%.(2)Constructing identification model of melon seed imaging data based on deep machine learning.Deep learning model based on Alex Net,VGG16,Res Net18,Res Net50 network models are used to identify the melon seed image data.For comparison,traditional machine leanring(support vector machine、linear discriminant analysis、Bayesian)model are also used to indentify the melon seeds based on the twelve contour features of melon seeds extracted by the Snake active contour image segmentation algorithm It is revealed that the accuracy rates of the training set and test set are 55% and 53% for the traditional machine leanring model based on Snake active contour image segmentation algorithm.The accuracy rates of the training set and testing set are respectively reach at 98% and 95% for the classification algorithm based on deep learning convolutional neural networks.(3)Constructing multi-category crop seeds identification model based on deep learing.An image data set of 40 kinds of crop seeds including corn and passion fruit is established.Based on this data set,identification model based on Alex Net,VGG16,Res Net18 and Res Net50 network models are constructed.The research results show that the accuracy rates of the training set of Alex Net and VGG16 are respectively 85.90% and 87.50%,and the accuracy rates of the test set are 85.86% and 86.5%.As far as the identification model based on both Res Net18 and Res Net50,the accuracy rates of both the training set and the test set reach a high value of 99%.In conclusion,hyperspectral imaging,machine learning and image processing technology are combined in this thesis to establish identification model of melon seed hyperspectral data based on machine learning,identification model of melon seed imgaging data based on deep machine learning and multi-category crop seeds identification model based on deep learing.The results revealed in this thesis has important significance for the identification of crop seed varieties.