Weak constraint leaf image recognition based on convolutional neural network

Abstract

Recently the computer vision and machine learning research communities pay a great attention to the leaf image recognition problem. Our literature survey focusing on the user interaction aspect reveals that two schemes of image acquisition have been used, one with strong constraint and the other with no constraint. The strong constraint interaction asks users to capture images by placing a leaf on a uniform background such as white paper while the unconstrained interaction allows any form of image capturing. The former one gets a high performance sacrificing the user convenience while the latter one provides a great convenience sacrificing the recognition performance. Our scheme is weakly constrained in the middle of two extremes. The proposed interaction scheme only asks users to center the leaf on smartphone camera screen. The leaf may be on the tree or off the tree. When the leaf is picked off the tree, it is recommended to place it against rather uniform background such as sky, soil, or tree bark. By fine-tuning the pre-trained CNNs (Convolutional Neural Network), we obtained a practical performance, 96.08% top-1 and 99.81% top-5 accuracies. The dataset is publicly open and the recognition system is released as an Android App.