Deep learning for automatic identification of plants through leaf

Automatic identification of plants, has been a widely explored field for the conservation of environment. Deep Learning has been extensively used in image recognition tasks due to its powerful ability to extract features from the given set of images. In this paper, we have trained Convolutional neural Network models from scratch by first pre-processing the images using MobileNet’s pre-processing input function to identify the plant species using leaf images. Four CNN models are discussed at different depths to understand how the accuracy of identification can be improved and the impact of hyperparameters namely batch size and number of epochs have on the accuracy of identification. The four models have been evaluated on two freely available leaf datasets: Flavia and Swedish. To reduce overfitting, data-augmentation and Early Stopping callback has been applied. The performance of the proposed CNN model was also compared to SVM, Random Forest and K-Nearest Neighbors classifiers on both datasets. Maximum accuracies were reported to be 95.35 % and 95.24% on Flavia and Swedish respectively.