Leaf Species Recognition Based On Deep Convolutional Neural Network

In nature,the distribution of plants is very extensive.The survival and life of human beings are inseparable from plants.The survival and development of other organisms on the earth also depend on plants.The identification ofleaf species is important for studying the evolution of plants,plant species protection,agriculture.Development and other aspects are very helpful.Due to the wide variety of leaves,the traits of some species are also similar,and the shooting of leaves is easily affected.The pictures have more noise interference,which makes the identification of leaf types more difficult.In recent years,with the rapid development of deep learning,convolutional nexural network technology has been widely used in various fields such as agriculture,industry,military,etc.,and has achieved significant improvement in image recognition and classification.Therefore,this paper uses the convolutional neural network method to identify the 2014 China-Europe Forest Dataset(MEW 2014)provided by the Image Processing Department of the ASCR Institute of Information Theory and Automation in the Czech Republic.The main research contents of this thesis are:(1)Perform multiple pre-processing on the leaf image in the dataset,including background whitening,background dirt removal,removal of excess leaves in the background,etc.,and then using the InceptionV3 network for pre-and post-pretreatment comparisons.(2)The leaf dataset was augmented,expanding from 1,073 images to the last 80,400 images without changing its biological characteristics,and then using the InceptionV3 network for comparison before and after dataset expansion.(3)Based on the InceptionV3 network,the deeper and more complex InceptionResNetV2 network is used for training,which further improves the recognition rate.Finally,this thesis compares and analyzes all the experimental results.There are 201 types of leaf datasets used in this paper.When the leaf dataset is not pre-processed,the recognition rate using InceptionV3 network is 85.54%;After multiple preprocessing steps.the same method recognition rate is 93.15%;then the leaf data set is expanded,and the InceptionV3 network recognition rate is 98.28%on the expanded data set;finally,the deeper and more complex InceptionResNetV2 network is used.The recognition rate reached 99.98%.Compared with the original papers of the leaf dataset,the number of leaves has increased by 48.and the recognition rate has increased by 11.07%.