Research On Few-shot Learning Classification Method Of Plant Leaves Based On Metric Learning

Plants are widely distributed in the natural environment and play an important role in the protection of the earth's ecosystem.In the past,in the research of plant identification,the visual features usually used the leaves,flowers,fruits,stems and other plant organs,among which the leaves of plants are the most representative.At present,with the development of image technology,convolutional neural network in deep learning,with its outstanding performance in the field of computer vision,has become the main means to solve such problems.However,the disadvantage of using deep learning method is also obvious.The premise of obtaining high classification accuracy is that the network has enough supervised learning samples.In most cases,researchers can only obtain a small number of samples,and the general deep learning neural network is extremely poor in the face of a small number of learning samples.Therefore,the concept of small sample learning is proposed,which aims to learn the classification methods of these samples from a small number of supervised samples.In this paper,a method based on metric learning is proposed to extract leaf features and classify leaves.The main contents of this paper are as follows:1.Firstly,based on the condition of few-shot learning,this paper analyzes and constructs the structure framework of leaf feature extractor and feature classifier respectively,and constructs the overall framework using the idea of two-way parallel convolutional neural network.Using the classic open leaf dataset: Flavia dataset,Swedish dataset and Leafsnap dataset to form the experimental dataset,after reorganization,screening and image preprocessing,the experiment was carried out.2.We use a variety of excellent convolutional neural network feature extractors,such as Google Net,Desen Net,Res Net and so on,to extract features,and design a combination based on the small sample classification framework.In this paper,we present the performance and network structure of different convolutional neural networks in this framework,and carry out the classification accuracy experiment in the dataset,and give the actual results Test results and analysis.3.A spatial structure optimizer is proposed in this paper for the purpose of reducing the classification accuracy due to the small number of samples in the training task.Experiments show that the classification accuracy can be improved by 2% when the optimizer is used.