基于改进ResNet18的胡麻干旱胁迫识别与分类研究

Identification and Classification of Flax Drought Stress Based on Improved ResNet18

【目的】为实现在移动端对胡麻干旱胁迫实时监测,解决传统机器学习方法在识别分类时准确率低、速度慢的问题,提出一种基于改进ResNet18的胡麻干旱胁迫分类识别方法。【方法】首先在网络中添加卷积块注意力(CBAM)模块,强化网络对胁迫特征的提取能力;其次调整残差块中批标准层、激活函数、卷积块的连接顺序,实现对输入的样本数据进行归一化操作;最后将ReLU激活函数替换成LeakyReLU激活函数,避免出现神经死亡现象。试验分为无胁迫、轻度干旱、重度干旱3个水分胁迫处理,分批次采集不同干旱程度胡麻叶片图像,数据样本按3∶1分为训练集与测试集,并使用数据增强的方法增加样本的多样性。【结果】改进ResNet18模型分类准确率高达98.67%,相比于ResNet18和VGG16分别提高6.14和4.87个百分点,而模型所需参数大小仅为42.80 MB,单幅图像推理时间为17.50 ms。【结论】该文模型对胡麻干旱胁迫具有更好的分类识别效果,能够实现嵌入式设备上胡麻干旱胁迫识别的实时性要求。可为胡麻干旱监测、机械化生产等研究提供技术支持。

[Objective]In order to realize the real-time monitoring of flax drought stress on the mobile terminal,and solve the problem of low accuracy and slow speed of traditional machine learning methods in recognizing and classifying,this study proposes a classification and recognition method of flax drought stress based on improved ResNet18.[Method]Firstly,the Convolutional Block Attention(CBAM)module is added to the network to strengthen the network’s ability to extract features;secondly,the connection order of the batch standard layer,activation function,and convolutional block in the residual block is adjusted to achieve the normalization operation of the input sample data;and lastly,the ReLU activation function is replaced by the LeakyReLU activation function to avoid the phenomenon of neural death.The experiment was divided into three water stress treatments,namely,coercionless,mild coercion and severe coercion,and the images of flax leaves with different drought degrees were collected in batches,the data samples were divided into the training set and the test set based on the proportion of 3∶1,and the method of data enhancement was used to increase the diversity of the samples.[Result]The test results show that the classification accuracy of the improved ResNet18 model is as high as 98.67%,which is 6.14 and 4.87 percentage points higher than that of ResNet18 and VGG16,respectively,while the required parameter size of the model is only 42.80 MB,and the inference time for a single image is 17.50 ms.[Conclusion]The model of this study has a better classification and recognition effect on flax drought stress.It can realize the real-time requirements of flax drought stress recognition on embedded devices,thus providing technical support for the research of flax drought monitoring,mechanized production and so on.