摘要
自然场景下采集的茶叶病害样本背景复杂,且存在类别样本数量不平衡的现象。结合茶叶病害特征,提出一种基于改进MobileNetV2的茶叶病害识别方法。在MobileNetV2倒残差结构中引入坐标注意力机制,使网络将注意力定位于目标区域,减少无关信息的干扰,有效地学习茶叶病害特征。同时,将交叉熵损失替换为焦点损失,解决茶叶病害样本类别不平衡导致网络训练效果不佳的问题。在茶叶病害数据集上进行验证实验,实验结果表明,改进后的MobileNetV2网络识别率达96.31%,参数量仅为2.27MB,对比其他模型具有较高性价比。改进后的MobileNetV2网络能高效地对自然环境中茶叶病害进行识别,为茶叶病害识别提供了新思路。
The background of tea disease samples collected in natural scenes is complex, and there is an im-balance in the number of categories of samples. Combined with the characteristics of tea diseases, a tea disease identification method based on improved MobileNet V2 is proposed. The coordinate attention mechanism is introduced into the MobileNet V2 inverted residual structure, so that the network can focus on the target area, reduce the interference of irrelevant information, and effec-tively learn the characteristics of tea diseases. At the same time, the cross-entropy loss is replaced by the focal loss to solve the problem of poor network training effect caused by the imbalance of tea disease sample categories. The verification experiment is carried out on the tea disease data set. The experimental results show that the improved MobileNet V2 network has a recognition rate of 96.31% and a parameter size of only 2.27 MB, which is more cost-effective than other models. The improved MobileNet V2 network can efficiently identify tea diseases in the natural environment, which provides a new idea for the identification of tea diseases.
出处
《软件工程与应用》
2022年第4期743-750,共8页
Software Engineering and Applications
