| Corn is one of the important food crops in our country,it is of great significance to detect the quality of corn seed objectively and efficiently,and to grasp the quality of corn timely and accurately for ensuring food safety.The quality of corn is different due to the damage caused by diseases,pests,mold and agricultural machinery in the process of growing,harvesting,storage and transportation.Traditional grain quality inspection relies heavily on manual inspection and uses chemical and physical methods,which is not only inefficient,but also causes food waste and environmental pollution.With the development of agricultural information technology,machine vision has the characteristics of objectivity,speed and efficiency compared with traditional detection methods.Many scholars have carried out the detection research of different agricultural product quality.In particular,with the development of deep learning in recent years,the research on automatic real-time recognition of agricultural product quality based on RGB images has achieved certain results.Aiming at the problem of corn seed quality classification and detection,this paper uses a variety of machine learning and deep learning models,uses CNN and Transformer structure for feature extraction,combines the local features and global features of corn seeds for collaborative reasoning,and constructs a corn seed classification model based on convolutional neural network,selfattention mechanism and support vector machine.The specific research content is as follows:(1)Research on maize seed quality recognition algorithm based on machine learning.In order to achieve objective and accurate classification of corn seed quality,this paper collected 2700 corn kernels of 9 kinds of different quality from 3 corn varieties as data samples,and extracted 41 features of corn seed color,shape and texture features.In order to further improve the recognition efficiency of corn seeds and reduce the amount of calculation,PCA is used to reduce the dimension of features and retain the first 12 principal components as the input of the model.Four machine learning models including KNN,SVM,RF and MLP are constructed,and the parameters of the model are optimized by cross-validation.The results show that the classification accuracy of the SVM model with RBF as the kernel function is up to 97.78%.(2)Research on corn grain quality algorithm based on deep learning.In order to make the convolutional neural network learn the characteristics of corn seeds more fully and comprehensively,the data augmentation is realized by rotation and adding noise,and a data set of 16200 corn seeds is constructed.Five convolutional neural network models including VGG16,Goolenet,Res Net50,Mobile V3 and Mobilevit are built,and different feature extraction structures such as multi-scale,residual,depthwise separable convolution,channel attention and Transformer are comprehensively compared.The characteristics of corn seed feature extraction.Among them,Res Net50 has the highest accuracy of 99.27%,Mobilenet V3 achieves the fastest detection rate of 0.01 S per graph,and the minimum model parameters is Mobilevit with only 0.95 M parameters and an accuracy of 99.20%.Using Mobilevit as a feature extractor and PCA for data dimensionality reduction,only the first 9 principal components are needed to achieve more than 95% of the cumulative contribution,and the classification accuracy is 99.63% when trained with SVM.Experiments confirm that the feature extraction structure combining depthwise separable convolution and Transformer is suitable for corn seed quality classification tasks. |
