Study On Wild Fungus Classification Model Based On Improved ShuffleNet V2 And Migration Learning

There are many types of wild mushrooms in China,and the classification and identification of wild mushrooms are closely related to their industrial development.Traditionally,when people encounter unknown wild mushrooms,they use the phenotype information to search for them in browsers,and the search results depend on the text information provided by the describers,which is subjective.With the development of technology,people can upload pictures to the internet to identify the species.In traditional identification methods,it is necessary to remove the background information and then extract features and train models,but the efficiency and accuracy of identification are not optimistic.Poisoning incidents from wild mushrooms frequently occur,especially in Yunnan and other places,which has become one of the obstacles to industrial development and the export of wild mushrooms.Therefore,finding an efficient wild mushroom identification method is particularly important.Deep learning technology has been widely used in the field of image recognition and has achieved good recognition results.This paper selects a lightweight convolutional neural network to train a wild mushroom dataset,and improves it to quickly and accurately identify the types of wild mushrooms.The main work of the paper is as follows:(1)Firstly,obtain the wild mushroom image dataset from the Paddle Paddle platform,clean the data,and further use data augmentation methods to expand the images of four types of wild mushrooms with small quantities to make the image data of each type of wild mushroom balanced.The expanded wild mushroom image dataset has 9 categories and a total of 5019 images.(2)Secondly,select three commonly used lightweight convolutional neural networks for experimental comparison,and select ShuffleNet V2 as the main model for improvement due to its high accuracy and fast convergence speed in the training set and test set.To address the issues of small sample datasets and complex backgrounds,ECA attention module and transfer learning are added to propose the TL-ShuffleNet V2-ECA network model structure,which enables the model to focus more on the features of wild mushrooms themselves,reducing the feature weights of the background area,and thus extracting the target features such as color and texture of wild mushrooms in a more fine-grained manner.(3)Finally,accuracy,loss,and model training time are selected as evaluation indicators for the model classification performance.The TL-ShuffleNet V2-ECA model proposed in this paper is compared and analyzed in depth with the original model,two other comparative models,and the classification performance is evaluated.The three evaluation indicators of the TL-ShuffleNet V2-ECA model are the highest.After optimization,the accuracy of the model is greatly improved with a small increase in training time(only 2 seconds per epoch).The training set accuracy reaches 96.6%,which is respectively 8.9%,8.7%,and 4.8% higher than the original model,Dense Net,and Efficient Net V2.The test set accuracy reaches 93.3%,which is respectively 9%,8.9%,and 9% higher than the original model,Dense Net,and Efficient Net V2.In order to analyze the contribution and effect of the ECA attention module and transfer learning in the model,the proposed complete model is subjected to ablation experiments.The experimental results show that the TL-ShuffleNet V2-ECA model has more excellent performance,which provides technical support for the subsequent research on wild mushroom and other image classification and recognition,and opens up new ideas.