Research And Application Of Butterfly Classification Algorithm Based On Depth Learning

Butterflies,as one of the common insects on Earth,have significant ecological importance and close connections with human production and livelihood.However,due to their diverse and complex morphologies,it is difficult for the general public to accurately identify different butterfly species,resulting in issues such as confusion between harmful and beneficial species,unclear species classification,ambiguous taxonomic groups,name confusion,and misidentifications.This situation seriously affects the conservation of rare and endangered butterfly species and the control of harmful species.Therefore,the development of an efficient and accurate butterfly classification algorithm that simplifies the identification process,produces accurate results,and is easy to use,is not only important for biodiversity conservation and the popularization of butterfly species knowledge in China,but also has practical applications in the development of butterfly resources.In this thesis,94 butterfly images of 6 families in China are collected and sorted,and the images are filtered and expanded,more than 100 samples of each species were collected,and a total of 6214 images of 50 species were collected as a data set.The main work of this thesis will be explained and analyzed through the following four aspects.(1)Firstly,to address the issue of diverse butterfly species that are difficult to distinguish,a comparison of deep learning image classification algorithms such as VGG,Transformer,Res Net,and Xception was conducted to analyze their differences in butterfly image classification performance.When building the butterfly classification model,training sets were constructed using butterfly specimen images from each species,and data augmentation techniques such as image rotation and noise addition were employed to expand the dataset,thus addressing the challenges of small-scale datasets and fine-grained classification of different species within the same genus or family.(2)Furthermore,the application of ensemble learning to improve classification performance.Due to the performance variability of the base classification models on the test set,the method of ensemble learning is applied to leverage the strengths of different models,such as VGG,Transformer,Res Net,and Xception,at the architectural level,to construct an integrated butterfly image classification model.Ensemble learning methods can effectively leverage the complementarity among different models to improve the overall classification performance.(3)In addition,the data set is further optimized through expansion.The Stable Diffusion method is used to expand the data set by adding the selected butterfly images to the training set,and the ensemble model is further trained using the expanded data set.This optimization at the data level aims to improve the classification performance of the ensemble model and enhance its generalization ability.(4)Finally,a online butterfly image classification system is constructed.The front-end of the system is designed using the Py Web IO framework,while the back-end is deployed and calls the optimized ensemble classification model using the Flask framework.The system allows users to upload butterfly images for classification and receive classification results,as well as providing information about butterfly species knowledge.This system provides a practical tool for users to classify butterfly images and access relevant butterfly species knowledge.