Improved Deep Learning-based Approach for Real-time Plant Species Recognition on the Farm

In this paper, a plant species recognition framework that combination of plant feature extractor and deep neural network is proposed. The morphology-based image processing technique is used to generate annotated feature of plant images, which are used to train deep classifier. In addition, geometrical transformation method is employed to augment the training data. Comprehensive experiments on training dataset with and without image pre-processing for plant species recognition are conducted to evaluate the performance of proposed approach. The results illustrate that the use of image pre-processing method can faster achieve a average loss than the method of not using pre-processing. Finally, the classifier utilizes images captured by various embedded cameras in the cultivation field and processes them through graphics processing units (GPUs) in real-time system. The experimental results demonstrate that the deep classifier can effectively recognize three plants, including Lollo Rosso lettuce, leaf lettuce and Djulis, which can apply to different scenarios around growth areas of plant in the cultivation field.