Classification and Variety Identification of Corn Ears using Machine Vision combined with Convolutional Neural Network

Abstract

Corn is an important human food crop and animal feed source. Purity of corn seed is critical to yield and marketing. Screening of corn ears is an important but time-consuming and labor-intensive task in seed production. In recent years, deep learning has made great achievements in tasks such as image classification, object detection, face recognition, etc. In this paper , a method combining convolutional neural network VGG16 and machine vision to quickly classify different varieties of corn using corn ear images is proposed. By collecting RGB images of corn ears with intact phenotypic traits of 5 varieties, a data set containing 1000 images was constructed, and divided into training set, validation set and test set according to the ratio of 7:2:1. By improving the fully connected layer structure of the VGG16 network, optimizing the training parameters, and using transfer learning and data enhancement techniques, the optimal performance model was obtained after training all layers of the VGG16, and the accuracy rate reached 98.00% on the test set. Under the same experimental conditions, comparing the three methods of training from scratch, pre-feature extraction and only training the fully connected layer, the accuracy rates obtained were 94.00%, 96.88%, and 94.00%, respectively. The improved model achieved the highest classification accuracy rate and stable performance. In the experiments, the effects of network parameters on the model classification results were also discussed. The experiment showed that the phenotypic characteristics of the corn ears could better realize the classification and identification of different varieties of corn, which provides a reference for the intelligent sorting of corn seed production.

Code and data availability

The preprint's Data Availability Statement deposits the paper's own corn ear image dataset (1000 RGB images of 5 varieties used for VGG16 classification) on Mendeley Data with a public DOI, making it a paper-specific, publicly actionable phenotyping image dataset.