LAUSR

Image identification technology has great significance for forestry production and forestry management. Highly similar object identification tasks, such as tree species with similar leaves, are extremely challenging. Simply using typical Convolutional Neural Networks (CNNs) or simply adding more convolutional layers still performs poorly in the above tasks. In this paper, we present a novel attention mechanism to enhance the CNN for identification of tree species with highly similar leaves. This paper presents a highly discriminative network, namely attention branch based convolutional neural networks (ABCNN), to better distinguish the differences between leaves features. Firstly, we proposed a novel structure, in which an attention branch is added in all block layers of network besides the typical normal branch. Secondly, our attention branch adopts a condensation process to obtain a region of interest (ROI) from global information of input and designs a reconstruction process to amplify the features difference to focus on the ROI. Thirdly, we design a fusion process, which carefully combines the attention branch with a normal branch to improve the network performance in the training process. The proposed ABCNN is tested on special dataset of Leafsnap with highly similar tree leaves. Our approach achieved 91.43% classification accuracy, which is higher than previous methods. Furthermore, ABCNN is also tested on general data set of SVHN and obtains 98.27% classification accuracy, which is the most competitive when considering the lower computational resources for ordinary applications. Both above experiments demonstrate the discrimination and robustness of the proposed method.