Application Research Of Deep Neural Networks In Plant Classification At Species And Variety Level

Deep Neural Network(DNN)explodes sharply in various research fields,such as detection,classification,recognition and supervision,due to the rapid expansion of computing resources and computing efficiency.Recognition and taxonomy of plants are vital in recognizing,understanding and protecting species diversity.In traditional plant classification,it's necessary to carry out long-time intensive training for professionals in plant taxonomy in advance to obtain sufficient experience,which makes it difficult for non-professionals.In the domain of visual cognition,the recognition of objects runs throughout the whole learning phase of human beings.While in biology,recognition and taxonomy become harder and harder because of the difference between species becoming smaller and smaller with the classification level from high to low.How to apply DL methods to plant species-below classification is an import task in current research.With the evolution of DL,it becomes more and more powerful and intelligent in classification.Mobile phone and terminal databases as well as other relevant technologies provide abundant image data,which in turn is accelerating the optimization of DL algorithms.This work reviewed the status quo application of DL in plant taxonomy and selected 6DL models which experimented on 5 plant datasets.Four of them are public datasets at species level that are Flavia,Swedish leaf,D-leaf and Oxford Flower102,respectively.The other one is a plant dataset at cultivar level named Camellia@clab which is photographed by ourselves.The camellia species belongs to the genus Camellia sasanqua Thunb.There are many artificially cultivated and wild developed cultivars below this species.Each cultivar has high similarity and is difficult to classify.We provided a new camellia cultivars dataset,and six models employed to train this dataset.The chosen 6models are Le Net,Inception Net,Res Net,Dense Net and Mobile Net.Finally,relative research prospects are outlooked after analyzing and comparing the outcomes of experiments.The outcomes concluded that the average taxonomy accuracy of models applied in our experiments was above 75% on all datasets.Dense Net achieved the best classification performance with classification accuracy over 93% on every open dataset,but it has the highest computational resources,and the longest calculating time is required.Le Net performed well on these datasets with classification accuracy over 82.23%,and the requirements of computing resources and calculating time were moderate.The results certified that performance of each model declined on Camellia@clab,because the high similarity among cultivar accessions aggravated classification difficulty,but Dense Net and Le Net achieved a nice classification result.On the other hand,loss of each model is reduced in Camellia@clab,which demonstrates that big data size is useful for convergence of DNN classification model.