Attention-Based Deep Multi-scale Network for Plant Leaf Recognition

Abstract

Plant leaf recognition is a computer vision task used to identify plant species. To address the problem that current plant leaf recognition algorithms have difficulty in recognizing fine-grained leaf classification between classes, this paper proposes a DMSNet (Deep Multi-Scale Network) model, a plant leaf classification algorithm based on multi-scale feature extraction. In order to improve the extraction ability of different fine-grained features of the model, the model is improved on the basis of Multi-scale Backbone Architecture model. In order to achieve better plant leaf classification, a visual attention mechanism module to DMSNet is added and ADMSNet (Attention-based Deep Multi-Scale Network), which makes the model focus more on the plant leaf itself, is proposed, essential features are enhanced, and useless features are suppressed. Experiments on real datasets show that the classification accuracy of the DMSNet model reaches 96.43%. In comparison, the accuracy of ADMSNet with the addition of the attention module reaches 97.39%, and the comparison experiments with ResNet-50, ResNext, Res2Net-50 and Res2Net-101 models on the same dataset show that DMSNet improved the accuracy by 4.6%, 18.57%, 3.72% and 3.84%, respectively. The experimental results confirm that the DMSNet and ADMSNet plant leaf recognition models constructed in this paper can accurately recognize plant leaves and have better performance than the traditional models.