Research And Implementation Of Rare And Endangered Plant Leaf Recognition Method Based On Transfer Learning

Plant leaves are an important basis for identifying plant species.Most of the traditional methods for identifying plants are manual identification,which not only has a large task and low efficiency,but also is difficult to ensure the accuracy of identifying plant data.With the rapid development of science and technology,it is no longer necessary to identify plants manually.Mobile software can help identify plants quickly.For rare and endangered plants,the sample size of leaf data set is small,Leaf recognition is mostly applied in the use scenarios with low computational power,such as intelligent mobile terminals.This experiment studies a recognition method of rare and endangered plant leaves based on transfer learning and knowledge distillation technology.The main experimental results include:(1)The data set sample database of 16 rare and endangered plant leaf species was established,and the automatic learning of the recognition model to the deep plant features of the leaf image was realized through the training of convolution neural network.(2)The migration learning technology has effectively solved the related problems of the convolution neural network over-fitting caused by the limited sample size of the rare and endangered plant leaves.At the same time,the migration model can make the deep features have a wide range of applicability,and can obtain better recognition accuracy than the direct training of the convolution neural network.(3)The lightweight recognition model of endangered plant leaves after knowledge distillation has been created,which is very close to the relative performance of the relatively complex migration learning network recognition model,while the scale of the parameter quantity is only 5.64%-19% of the scale of the complex target detection network parameter quantity,significantly reducing the complexity of the network model and improving the operation performance of the network model,The application scenario of the rare and endangered plant leaf recognition model is expanded.(4)The Android APP for leaf identification based on Tensorflow Lite framework and the corresponding server interface have been developed,realizing the functions of login registration,leaf identification,rare plant classification,rare plant recommendation,discovery,collection,search,sharing and uploading,leaf collection,plant browsing history,tree query platform,etc.Discussion and analysis of experimental results: First,a dataset sample database of rare and endangered plant leaf species was created.Second,replace the extended data set of the blade with the trained Alexnet The last layer in VGG16,Goog Le Net and Res Net models,namely the full connection layer,is followed by migration learning,which significantly improves the recognition accuracy of rare and endangered plant leaf images.Thirdly,the knowledge distillation technology is used to build a lightweight model based on the knowledge of Alexnet,VGG16,Gog Le Net and Res Net models.Fourth,compared with other plant leaf recognition methods,the application of the light-weight rare and endangered plant leaf recognition model not only has a higher accuracy of rare and endangered plant leaf recognition,but also reduces the complexity of the recognition model.The leaf recognition model is constructed by using the convolution neural network technology,and the size of the model is compressed under the premise of maintaining a high recognition rate,After the model is transformed into Tensorflow Lite model,it is deployed through Android APP mode to realize automatic identification of leaf types.Fifthly,the experimental research uses Java programming language to store the image information of all endangered plant leaves in the My SQL database of the network server,which can not only effectively reduce the workload of relevant personnel,but also improve the accuracy of plant species classification.