Identification Of Amanita By Deep Learning Model

The Amanita mushroom,known as the "angel of destruction",has a faint fragrance and is often mistaken for non-toxic mushrooms,leading to a high mortality rate from poisoning.However,it is widely used in food processing and medical fields.Amanita mushrooms come in different shapes and colors,and most of them can be identified by their appearance.Traditional methods for identifying them in the wild or during factory sorting are not comprehensive,efficient,or accurate enough.Therefore,this study focuses on using deep learning models to classify Amanita mushrooms and detect their growth status based on their external features,such as cap color,cap surface shape,and overall shape.The main research content and conclusions are as follows:(1)This study collected images of different types of Amanita mushrooms,covering different brightness,angles,and sizes,to construct a complete Amanita mushroom image dataset and performed classification.Given that different growth environments can lead to differences in Amanita mushroom image data,data augmentation methods such as translation,rotation,noise addition,and cropping were used to expand the original dataset to about five times its size,improving the quantity and quality of the dataset to enhance the model’s generalization ability,robustness,and training efficiency,which has practical application value.(2)This study compared the performance of five mainstream detection models in terms of Amanita mushroom loss functions and accuracy and found that the YOLOv5 s algorithm has advantages such as low complexity,fast detection speed,and high detection accuracy.Therefore,this study proposes an Amanita mushroom recognition model based on YOLOv5 s and analyzes the model using evaluation indicators.The results show that the YOLOv5 s model has good convergence and classification detection performance for different Amanita mushrooms,with an accuracy of 96.4%.(3)To improve the target detection performance of the model,this study introduced the Coordinate Attention(CA)mechanism into the YOLOv5 model and used Concat feature fusion to enhance feature extraction ability,and finally used the DIo U＿NMS loss function to improve accuracy.After improvement,the accuracy was98.6%,a 2.2% increase compared to before.(4)To reduce the complexity of the algorithm and the memory and computational requirements of the deployment device,this study designed a lightweight structure for the YOLOv5 s algorithm and compared the performance of three lightweight networks,Ghostnet-YOLOv5 s,Mobile Netv3-YOLOv5 s,and Shufflenetv2-YOLOv5 s.GhostnetYOLOv5s had the best detection accuracy on the Amanita mushroom test set.Compared to YOLOv5 s,the parameter volume was reduced by about 50%,FLOPs decreased by 10.5G,and the model size decreased by 7.41 M.