Large convolutional neural networks are difficult to deploy in practical applications because of the complexity of models,while lightweight networks are often less accurate than the former because of the optimization of model structure.To solve these problems,ShuffleNet V2 was improved and a lightweight MAM-ShuffleNet citrus disease recognition model was proposed.Firstly,the mixed attention module(MAM)was introduced in ShuffleNet V2 to improve the ability of the model to extract disease features.Secondly,Ghost module was used to optimize the convolutional layer in the network,which effectively reduced the number of network model parameters and calculation cost.Finally,the stacking times of ShuffleNet V2 unit in the network structure were adjusted to further simplify the network parameters.The results showed that the average recognition accuracy of MAM-ShuffleNet model reached 97.7%in the self-built citrus leaf data.Compared with the original ShuffleNet V2,the number of parameters was reduced by 45.7%,and the recognition accuracy was increased by 1.2 percentage points.The comprehensive performance was better than ResNet50 and DenseNet121 models.
维普
万方数据