Research On Corn Seed Variety Identification Based On Deep Learning And Machine Vision

Corn is one of the major agricultural crops in China,and plays a very important role in the development of our national economy and social stability.While corn seed is the upstream material in the process of corn production,the purity of the seed will directly affect the quality of corn.Therefore,the accurate identification of corn seed varieties is of great importance for agricultural production.In this paper,a deep learning approach is used to classify different varieties of corn seeds.First,a standard dataset with 32,500 maize seeds containing 19 different varieties is established,and two improved models are proposed to identify the varieties of maize seeds from two perspectives to achieve accurate and efficient classification of maize seeds.The details are as follows:(1)To solve the problems of strong subjectivity,low efficiency and low accuracy of traditional maize seed recognition methods,this paper proposes a maize seed recognition classification method based on multi-scale features and feature attention fusion based on Swin Transformer.Since the shallow network contains more geometric detail information,it is able to extract more seed key minutiae features.Therefore,this paper considers retaining both shallow and deep network features to obtain richer maize seed features;at the same time,introducing feature attention layer to give different weights to features with different degrees of importance;finally fusing shallow and deep features to construct multi-scale fusion features of maize seed images and classifying seed images into 19 varieties by classifier.The results show that the accuracy of the MFSwin Transformer model reaches 96.47%,which is 3.56% more accurate than the original model,while the parameter memory is 12.83 M,which is about half of the original model.The model is able to obtain richer seed features and has high recognition performance.(2)To solve the problems of traditional convolutional neural networks in corn seed recognition with huge parameters,long training time,and difficulty in deploying to mobile devices for application,this paper proposes a classification method for corn seed recognition based on Mix Conv and Triplet Attention on the basis of Mobile Net V3.The Mix Conv structure is introduced after the first layer of the original model,which mixes several convolutional kernels of different sizes to facilitate the acquisition of more detailed features from seed images;the SE module in the original model is replaced with the Triplet Attention mechanism to reduce the parameters of the model while improving the model accuracy.The results show that the accuracy of MT-Mobile Net V3 model reaches 94.95%,which is 3.56% better than the original model,and the parameter memory is 2.736 M.Compared with other lightweight networks such as Efficient Net and Ghost Net,MT-Mobile Net V3 model has obvious advantages in recognition accuracy,parameter memory and recognition speed.Compared with other lightweight networks such as Efficient Net and Ghost Net,the MT-Mobile Net V3 model has obvious advantages in terms of recognition accuracy,parameter memory and recognition speed,and can meet the requirements of corn seed variety recognition model deployed in various mobile and embedded devices.In summary,the two models proposed in this study are able to achieve higher accuracy of maize seed image recognition,while effectively reducing the number of parameters of the model and providing ideas and references for seed variety recognition.