At present, deep neural network models are widely used in wild mushroom identification tasks and have achieved good results. However, these neural network models still have the problem of large parameters and difficulty in deploying on mobile terminals. To address this problem, this paper proposes a BE—EfficientNet method based on the fusion of dual attention mechanisms. This method uses the lightweight model EfficientNetB0 as the benchmark model. First, the dual attention module BE, which is a fusion of BAM and ECA, replaces SENet in the core module MBConv of EfficientNetB0, so that the model not only obtains channel feature information, but also spatial feature information; at the same time, Adam optimization is introduced. The device realizes adaptive adjustment of learning rate and improves classification accuracy. Experimental results show that the improved BE—EfficientNet model is 2.9% more accurate than the baseline model EfficientNetB0, with a parameter size of 4.40 MiB. In addition, the proposed fused dual attention mechanism BE was applied to VGG16, ResNet50, MobileNet V2, GoogLeNet and ShuffleNet V2 models for wild bacteria identification, and the accuracy was improved by 0.5%, 0.8%, 0.6%, 0.5% and 1.0%, respectively,  indicating that the dual attention mechanism BE has certain versatility. This method can provide a new solution for deploying wild mushroom identification on the mobile terminal.