Application Research Of ResNet And VAE Model In Image Recognition Of Large-flowered Chrysanthemum Cultivars

Chrysanthemum × morifolium Ramat.is a world-famous ornamental plant that is widely cultivated in the world.Large-flowered chrysanthemum cultivar is a potted group among them,and since its spread,it contains rich traditional Chinese culture and unique cultivars.Due to the abundant flower types of these large-flowered chrysanthemum cultivars,it is difficult to achieve accurate and efficient variety testing and identification only by referring to the "DUS Testing Guide for New Plant Cultivars-Chrysanthemum".In recent years,image processing technology based on machine vision and deep learning models based on data analysis have provided new methods for plant image recognition and character extraction.In this study,the image collection device designed by the research group was used to collect images of Chinese large-flowered chrysanthemum cultivars for two consecutive years,and large-flowered chrysanthemum cultivar image database was constructed.Based on ResNet(Residual Neural Network)and VAE(Variational autoencoder)models,the large-flowered chrysanthemum cultivar images were captured.Recognition research aims to build a high-precision chrysanthemum cultivar identification and classification model to provide new tools for the identification and protection of chrysanthemum cultivars.The main results obtained in this study are as follows.1.A total of 213 cultivars were selected according to the Chinese traditional chrysanthemum cultivar classification criteria.They were cultivated and maintained in 2018 and 2019 respectively.The chrysanthemum cultivar image collection device was used to collect the cultivar images.A total of more than 16,000 images were obtained.Finally,through cultivar labeling and coding of flower and petal information,image datasets of 126 cultivars in 2018 and 118 varieties in 2019 was established.At the same time,the data is enhanced by zooming and cropping,which increases the number of pictures by 10 times.2.Based on the top-view images of the 2018 image dataset,a preliminary construction of the chrysanthemum cultivars Variational Autoencoder(VAE)network was found: The chrysanthemum cultivars image generated by the VAE can reflect the general characteristics of chrysanthemum cultivar,the distribution of ray and disc flowers,the shades and gradients of the chrysanthemum cultivars color.The use of VAE proves that generative learning is a new method for chrysanthemum research in the future.It also shows that the injection of more botanical information is extremely important to improve the accuracy of the chrysanthemum cultivars identification model.3.Based on the idea of transfer learning,a pre-trained ResNet50 model was constructed.The research results show that the pre-trained model based on ImageNet ignores the color information of the chrysanthemum cultivars,so the method of using transfer learning has limitations for the identification of chrysanthemum cultivars with rich color characteristics.4.For the generalization study of non-pre-trained models,we discussed the composition of datasets,model selection,training methods and processes,etc.,and established the image recognition and classification model of chrysanthemum cultivars.The results show that the ResNet18 non-pretrained model with the poly-decay strategy has the strongest generalization ability and the highest test accuracy for the 2019 chrysanthemum cultivars recognition(Top-5 is 69.53%).The choice of sample size,model size and learning rate adjustment method of the dataset have a great influence on the generalization performance of the model.This model has poor generalization performance for the chrysanthemum cultivars with large morphological differences in different periods,the generalization performance is weak,and the deep network model can be further optimized by incorporating more chrysanthemum phenotype information into the network.5.Using AP clustering and hierarchical clustering to analyze the 512-dimensional depth features extracted by the ResNet18 non-pre-trained model,it was found that the model reflects the characteristics of clustering by color features,and divide the 126 cultivars in the 2018 dataset into 17 clusters.Through the principal component analysis 2019 dataset,it was found that the first 21 principal components explained the data to 80%,and the visualization results of the first two principal components showed that there was some aggregation of depth features in flower color,petal type,and flower type.In this study,deep learning is used to explore the identification and classification of traditional Chinese chrysanthemum cultivar images.Through testing,the ResNet18 initialization model of chrysanthemum cultivar image is proposed to be the best network model currently,the verification accuracy of Top-1reaches more than 70%,and put forward a new idea of integrating chrysanthemum botanical information into the deep network model.A 17-category classification system for the deep features of the chrysanthemum cultivar images was established by cluster analysis.The characteristic classification system provides new methods and data for the identification and protection of chrysanthemum cultivars,and optimizes the traditional classification system of chrysanthemum cultivars.