Classification of Camellia oleifera based on Enhanced VGG16 network

Abstract: Camellia oleifera is one of the four largest woody oil plants in the world. There are the largest planting area and yield among all woody oil plants, most of which are distributed in more than 1 100 counties and cities in 18 provinces of southern China. Various kinds of Camellia oleifera cultivars have emerged in the market, particularly with the continuous development of Camellia oleifera industry in recent years. Therefore, a rapid and accurate classification can be urgent to identify the different varieties of Camellia oleifera. Deep learning can also be expected to serve as a promising way for classification, due mainly to the strong performance in many operations, such as classification, detection, and segmentation. Although the VGG and ResNet network models have been commonly used for the classification in agriculture, some limitations still remain, such as too large weight space and low accuracy. In this study, a systematic classification was performed on the image of Camellia oleifera plant leaf using the enhanced VGG16 network. Four cultivars of Camellia oleifera were selected to test, including the Changlin No. 3, No. 4, No. 40, and No. 53 taken from the National Camellia oleifera Seed Base of Dongfanghong Forest Farm, Jinhua City, Zhejiang Province, China. An Epson Perfection V30 scanner was used to collect the data set of the Camellia oleifera leaf images with a clear texture. 1800 images were obtained for the four Camellia oleifera varieties each. The training set and test set were then divided into the proportion of 3:2, where the training set was 4 320, and the test set was 2 880. Some operations of data enhancement were performed on the image during training, such as brightness adjustment, and random enhancement. The Enhanced VGG16 network was constructed using the hard-Swish and ReLU6 activation function, Residual block, Dropout, L2 regularization, the inter-layer deletion, and structural adjustment for the VGG16 network model. The performance of Enhanced VGG16 network model was then evaluated to compare with the classical convolutional neural networks (AlexNet, VGG16, Resnet50, InceptionV3, Xception). The model was also trained and tested on the sampled Camellia oleifera data set. More importantly, the hyperparameters dominated the model training and performance. The test optimizer was selected as the RMSprop optimization, while the loss function was categorical_crossentropy, as well as the Batch Size and the learning rate were 4, and 0.000 01, respectively. Furthermore, the learning rate was reduced to 10% of the original, if the accuracy of two epochs training sets remained the constant, as the number of iterations increased. The epochs were set to 100, while the training stopped in advance, if the accuracy of the four epochs training set remained. The results show that the accuracy of the training and test set of the Enhanced VGG16 network model were 98.98% and 98.44%, respectively. The average detection time of a single image was 55.32 ms, and the space of weight was 90.6 MB. The accuracies of validation and test set were improved by 3.08 and 2.05 percentage points, respectively, compared with the original. The space of weight and the average detection time were reduced by 165.4 MB and 2.18 ms, respectively, indicating the better performance of the enhanced model. Additionally, the enhanced VGG16 network model performed better in the classification of the Camellia oleifera leaf, and was much easier to deploy in the mobile terminals and embedded devices, compared with AlexNet, VGG16, Resnet50, InceptionV3, and Xception networks. The Camellia oleifera leaf data set can be further expanded to enhance the images at different growth stages, in order to overcome the interference of light in the unstructured and field environment. This finding can also provide promising technical support for crop species identification.