Recognition of carrot appearance quality based on deep feature and support vector machine

Abstract

Carrot appearance quality is an essential indicator for carrot sorter, so it is crucial to identify carrot appearance quality in carrot processing accurately. This paper proposed an innovative carrot appearance detection method, using a convolutional neural network to extract image feature information, and a support vector machine for classification. Firstly, the deep features based on 12 models were utilized for training the support vector machine. The deep features of the three-layer full connection layer of the network models (AlexNet, VGG16, VGG19) were extracted and imported into the support vector machine. The results showed that: (1) The accuracy of deep features with SVM were superior to the transfer learning models, and the average accuracy had increased by 1.42%. (2) The accuracy of the three models (AlexNet, VGG16, VGG19) based on the initial full connection layer was 98.13%, 98.06%, and 97.88%, respectively. The best model was ResNet101 + SVM, which had a recognition accuracy of 98.17%. So, this method has positive significance for the development of carrot sorter.