A Convolutional Neural Network for Leaves Recognition Using Data Augmentation

Recently, convolutional neural networks (ConvNets) have achieved marvellous results in different field of recognition, especially in computer vision. In this paper, a seven-layer ConvNet using data augmentation is proposed for leaves recognition. First, we implement multiform transformations (e.g., rotation and translation etc.) to enlarge the dataset without changing their labels. This novel technique recently makes tremendous contribution to the performance of ConvNets as it is able to reduce the over-fitting degree and enhance the generalization ability of the ConvNet. Moreover, in order to get the shapes of leaves, we sharpen all the images with a random parameter. This method is similar to the edge detection, which has been proved useful in the image classification. Then we train a deep convolutional neural network to classify the augmented leaves data with three groups of test set and finally find that the method is quite feasible and effective. The accuracy achieved by our algorithm outperforms other methods for supervised learning on the popular leaf dataset Flavia.