Video Image Sequence Analyzing-based Plant Species Identification System

As a major forestry country in the world,China has extremely rich plant resources.In order to better study plants,it is important to study the classification of plants.Plant species identification plays a very important role in identifying new species,maintaining ecosystem balance,and developing productivity.Deep learning-based plant images classification methods do not need to manually extract image features.Through large-scale sample training,automatic learning of features is achieved,reducing the workload and improving the efficiency of plant species identification.This paper takes plant video image sequences as the research object,and uses deep learning methods to learn features from plant samples,thereby automatically extracting plant features(shape,texture,color,etc.),and training a plant species classification neural network model.Then,a series of video key frames containing plant characteristics are extracted through the network model,so that plants in a video containing multiple plant information are identified and plant species images information are extracted.First,construct a rich data set using image enhancement methods such as random cropping,mirror flipping,and random brightness changing.And the data is de-averaged and normalized as the input of the neural network to prevent the network from over-learning random noise,improve the robustness of the model,and give full play to the advantages of deep learning.In addition,MobileNet is used for transfer learning.Then the global average pooling layer replaces the fully connected layer,and a new 512-dimensional fully connected layer is added to improve the expressive ability of the network,shorten the learning process and increase the availability of the model,which is important for real-time classification on the mobile.Finally,the algorithm model is transplanted to the mobile terminal to design a fast and effective plant species identification system,and the recognition accuracy is tested in application scenarios.When samples are added during model training,image enhancement is performed on the samples at random,in order that learning efficiency can be improved.It also uses the loss function weighting method to solve the problem of sample imbalance.In the end,this method was tested through videos of 400 domestic common plants,and the recognition accuracy rate reached 96.2%,which proved the effectiveness of the algorithm.This system provides a fast and effective plant recognition tool for plant enthusiasts and researchers,and brings educational software that can improve plant species recognition ability for teenager,strengthening people's awareness of protecting plants and having good practical value.