Wild mushroom classification based on improved MobileViT_v2

Accidental ingestion of toxic or inedible wild mushrooms is very dangerous. Because identifying their species is a fine-grained classification problem, it is difficult to directly apply general deep learning classification methods to obtain excellent fine-grained classification results. For mobile application scenarios, MobileViT_v2 with characteristics of convolutional ViT blocks is selected as the basic network. In addition, lightweight coordinate attention (CA) blocks are inserted to enhance the ability of filtering local foreground areas and feature channels of the network. After the shallow local fine-grained features processed by SoftPool, skipping connections are introduced between blocks to achieve fusion with global semantic features of deep layers to help achieve fine-grained classification that pays more attention to local detail features. Experimental results show that the improved method can improve the classification accuracy of Top1 from 96.83% to 97.39%, and reduce the loss by about 10%, indicating stronger generalization ability. Moreover, our improved network outperforms typical lightweight and heavyweight ones (MobileNet_v3, ShuffleNet_v2, ResNet50 and ResNeXt50-32×4d etc.).