Research And Application Of Flower And Disease Identification Based On Residual Network

With the popularity of greenhouses,the types of flowers planted are increasing,so the process of flower classification and growth monitoring requires a lot of manpower and resources.And because of the many types of flowers,the small differences between classes,the large intra-class differences,and the complex environment,traditional image classification methods cannot solve such problems well.In view of the above problems,the use of deep learning to automatically classify flower images and identify leaf diseases has become a research hotspot,which has important application value in the intelligent management of garden flower gardens.Therefore,on the basis of indepth analysis of domestic and foreign research results,this paper uses deep learning and transfer learning as the theoretical guidance to conduct research on flower image classification and disease recognition.The main tasks are as follows:(1)Traditional flower image classification is based on manual manual selection of single features or multi-feature fusion and reclassification.This method generally has the disadvantages of low accuracy,high cost,and weak generalization ability.Aiming at this problem,a flower image classification method combining attention mechanism with ResNet18 was proposed.First of all,based on ResNet18 as the basic model,the idea of fully convolutional structure is applied to the network model,and the fully connected layer of ResNet18 is replaced by a convolutional layer to optimize the network model.Secondly,the hybrid domain attention mechanism is integrated into the optimized ResNet18.Finally,it is classified by the softmax layer.The experiments on Oxford17 flowers and Oxford102 flowers datasets show that the improved method has achieved good results.(2)Aiming at the problem of flower image disease recognition,this paper selects the azalea leaf disease image as the research object.The traditional model and depth model were used for comparative analysis experiments,and a method suitable for image recognition of azalea leaf diseases was studied.Because the traditional Bo F lacks spatial information,this paper uses an improved Bo F model for leaf disease image recognition.The algorithm incorporates ordered spatial information into Lab-based color features to form new spatial color aggregation features.The multi-class feature learning algorithm combines color features and local features extracted by SURF algorithm to achieve image classification.Improvements based on Lab color aggregation features The recognition rate of the Bo F model reached 89%.Due to the complex structure of the deep model,large parameter scale and easy overfitting,a disease image recognition method based on deep model transfer learning is proposed.The weight parameters were saved through pre-training on the flower image data set,and then the model was re-trained on the leaf disease data set.The experimental results reached a 96.87% accuracy rate,indicating that the method of transfer learning has good accuracy and Generalization performance.(3)According to actual needs,design and implement a prototype system for flower image classification.Export the flower classification model and disease recognition model into a folder.The user can select the pictures in the folder or the camera to perform real-time classification.At the same time,the user needs to select the corresponding model,and finally the classification results are given on the page.