To improve the accuracy of plant leaf image recognition with a small dataset of plant leaves, a convolution neural network (CNN) plant leaf image recognition method based on transfer learning is proposed. First, a plant leaf image database was expanded by pre-processing the original plant leaf images through random horizontal and vertical rotation and random zooming. The expanded dataset was then processed by mean removal and divided into training and testing sets at a ratio of 4:1. Second, transfer learning training was performed on the plant leaf dataset using existing models (AlexNet and InceptionV3) that were pre-trained on a large dataset. To ensure these models can be adapted to image recognition for plant leaves, the original parameters of the last fully connected layer were replaced, whereas those of all other convolution layers were retained. Finally, the method proposed in this paper was compared to support vector machine, deep belief network, and CNN through testing on the ICL database. A Tensorflow training network model was used in the comparison test, and the results were visualized by Tensorboard. The testing results showed a considerable improvement in recognition accuracy when using the pre-trained AlexNet and InceptionV3 models, where the training dataset accuracies were 95.31% and 95.4%, respectively.
Recognition of Voltage Sag Sources Based on Phase Space Reconstruction and Improved VGG Transfer Learning
