Research On Plant Identification Methods Based On Attention Mechanism

Plants are important components of natural world and the key to harmonious coexistence between human and nature.Plant identification plays a crucial role in the development of plant resources and the protection of ecological environment.Traditional plant identification methods have poor generalization performance and greatest limitations.This paper aims to achieve high accuracy of plant identification,and the main innovations are as follows:(1)The multi-scale plant dataset LZU200 is collected and organized in this paper.LZU200 mainly contains herbaceous plants common in northwest China,with a total of 116,171 images in 200 categories,and 202 1021 images in each category.LZU200 contains plant images at different growth stages,and the shooting angle,illumination and scale of the images change greatly,which leads to high requirements for plant identification methods.(2)The global position based circular convolution and channel attention based convolutional neural network(GPC-CANet)is proposed for plant identification.GPCCANet uses Res Net50 as the backbone,and utilizes horizontal and vertical circular convolution blocks to extract global information.This model utilizes position coding to preserve the position relationship between image pixels,and introduces channel attention to learn the weights of different channels during training process.In this way,the model can distinct between primary and secondary features.At the same time,GPCCANet uses group convolution to extend the channel dimension without increasing the number of parameters.Experimental results on LZU200,Wild Edible Plants and BJFU show that GPC-CANet has high accuracy and strong generalization ability.(3)The lightweight adaptive spatial and channel attention based multi-scale context convolutional neural network(SC-MCNet)is proposed for plant identification.In order to extract information from key parts of different plant images such as flowers and leaves,plant identification methods need to establish multi-scale receptive fields.Traditional convolutional neural networks only expand receptive fields by continuously increasing the network depth,which is prone to cause feature redundancy.However,SC-MCNet establishes receptive fields in different scales by setting and applying atrous convolution with different magnification and convolution kernel sizes.In this way,the model obtains multi-scale context information and prevents redundant features.At the same time,SC-MCNet introduces adaptive channel and spatial attention mechanism to extract important information from key parts of plants.In addition,SC-MCNet designs an appropriate model size and structure,and uses depthwise separable convolution to reduce parameters,resulting in a final parameter amount of only 5.15 M.Experimental results show that compared to popular lightweight models,SC-MCNet has low computational complexity and high identification accuracy.