Plant Identification Based on Multi-branch Convolutional Neural Network with Attention

Abstract

The identification of plants has great significance in the study of plants, and it has applications in plant classification and medical research. With the development of computer vision, plants identification based on deep learning methods can effectively carry out. At present, most of the existing methods use the traditional features such as the shape and texture features for plant identification, and these methods are applicable to plants with large differences in their shapes. Therefore, we develop a multi-branch convolutional neural network with attention (MCNNA) to extract the effective features, and the first part of MCNNA is an attention block, which is used to reduce the influence of background, and the latter part is multi-branch convolutional neural network, which is used to extract the multi-view features through multi-channel. Experiments have shown that our proposed method has better performances for the classification of plants than the traditional methods. We tested our method on the dataset BJFU100 and obtained final accuracy of 97.89%, and we have the accuracy of 93.35% on our own image dataset.