Phalaenopsis growth phase classification using convolutional neural network

During phalaenopsis production and sales, a large number of manual classification operations must be done because phalaenopsis in different growth phases have different sales grades and prices. These manual classification operations are time-consuming and labor-intensive, so there is an urgent need to develop a practical approach to automatically classify phalaenopsis by growth phases for industrial applications. To improve agricultural productivity and meet the demand of mobile vision applications, a lightweight classification neural network model based on MobileNets is proposed to automatically classify different growth phases of phalaenopsis on the RGB image dataset. Due to the limited amount of experimental phalaenopsis images got by manual method, transfer learning and data augmentation are introduced in this study to expand the volume of the dataset and enhance the robustness of the proposed model. By transfer learning, the proposed model is pre-trained on ImageNet dataset, and then trained and tested on the experimental phalaenopsis dataset. With batch stochastic gradient descent algorithm, the model is trained on the experimental dataset, the classification accuracy is 98.9%. To check if over-fitting exists or not, the production dataset is input into the model, the classification accuracy is 98.1%, over-fitting can be approximately considered as not occurred. For comparison, several popular convolutional neural network models were tested on the experimental phalaenopsis dataset, the results show that the proposed model has better computation performance and higher adaptability than those models on phalaenopsis classification by growth phase, and the model can provide the research basis for more sophisticated flower classification tasks in the future.