基于改进ShuffleNet V2的柑橘病害识别研究

Identification of Citrus Diseases Based on Improved ShuffleNet V2

大型卷积神经网络因模型复杂难以部署于实际应用,而轻量级网络因优化模型结构而导致精度往往不如前者理想。针对上述问题,对ShuffleNet V2进行改进,提出一种轻量化MAM-ShuffleNet柑橘病害识别模型。首先,在ShuffleNet V2中引入混合注意力模块(Mixed attention module,MAM),提升模型对病害特征提取能力。其次,利用Ghost模块优化网络中卷积层,有效降低网络模型参数量和计算成本。最后,调整网络结构中ShuffleNet V2单元堆叠次数,进一步简化网络参数。结果表明,在自建柑橘叶片数据集中,MAM-ShuffleNet模型平均识别准确率达到97.7%;与原始ShuffleNet V2相比,其参数量降低了45.7%,识别准确率提升了1.2百分点;综合性能明显优于ResNet50、DenseNet121等模型。

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.