| 吴鸿飞 1, 刘万学 2*, 冼晓青 2, 赵梦欣 2, 姚 青 1.基于改进MobileNet的轻量级外来入侵植物识别模型研究[J].植物保护,2024,50(1):85-96. |
| 基于改进MobileNet的轻量级外来入侵植物识别模型研究 |
| A lightweight identification model for alien invasive plants based on improved MobileNet |
| 投稿时间：2022-11-10 修订日期：2023-02-13 |
| DOI：10.16688/j.zwbh.2022710 |
| 中文关键词: 外来入侵植物 智能识别 通道剪枝 知识蒸馏 注意力机制 MobileNet |
| 英文关键词:alien invasive plants intelligent identification channel pruning knowledge distillation attention mechanism MobileNet |
| 基金项目:国家重点研发计划(2021YFC2600400) |
|
| 摘要点击次数: 135 |
| 全文下载次数: 111 |
| 中文摘要: |
| 外来入侵植物防治的首要任务是准确识别入侵植物种类, 然而外来入侵植物种类繁多, 存在类间同质和类内异质现象, 给技术人员甄别与防治外来入侵植物带来了挑战。为了准确、实时和高效地识别外来入侵植物, 本文提出基于改进MobileNet的轻量级外来入侵植物图像识别模型(MobileNet-LW)。以专业人员鉴定得到的113种 11 628幅外来入侵植物图像作为研究对象, 并按照6∶2∶2的比例划分训练集、验证集和测试集。通过Retinex、旋转和高斯噪声等方法对图像进行数据增强。为了减少类间同质现象对模型误检的影响, 在模型MobileNet基础上添加了SE通道注意力机制和深度连接注意力网络, 提高网络对关键特征的提取能力。为了降低模型计算消耗和内存消耗, 采用通道剪枝方法对网络瘦身;为了弥补剪枝后造成模型准确率降低, 采用教师网络-助教网络-学生网络的形式对剪枝后的网络进行知识蒸馏, 学生网络通过软知识的学习来提高识别外来入侵植物的准确率。通过消融试验测试模型的性能, 利用平均准确率、平均召回率和平均F1值3个评价指标, 对现阶段经典模型与改进后模型MobileNet-LW所获得的识别结果进行评价。消融试验结果显示, 在相同数据集条件下, 所有改进点对模型的性能都有所提升, 且改进后算法在外来入侵植物图像识别中准确率提高了5.4百分点, 模型参数量减少了约53%;模型对比试验表明, EfficentNet、DBTNet、ResNet-101、ConvNext和MobileNet-LW 5个模型平均准确率分别为72.3%、74.9%、76.1%、79.7%和86.1%, 表明改进后的网络提高了外来入侵植物的识别准确率。基于改进MobileNet的轻量级外来入侵植物识别模型对113种外来入侵植物识别具有较高的准确率, 且模型具有轻量化特点。 |
| 英文摘要: |
| The primary task in controlling alien invasive plants is to accurately identify the species of invasive plants. However, there are a wide variety of alien invasive plants, and some of them present the inter-class homogeneity or intra-class heterogeneity, which brings challenges to the identification and control of alien invasive plants. In order to identify alien invasive plants accurately, efficiently and in real time, a lightweight identification model based on the improved MoblileNet (MobileNet-LW) was proposed. The 11 628 images of 113 species of alien invasive plants identified by technicians were divided into the training set, verification set and testing set in a ratio of 6∶2∶2. The image data were enhanced by Retinex, rotating image and Gaussion noise. In order to reduce the false detection, the SE channel attention mechanism and deep connection attention network were added to the MobileNet model to improve the ability of key feature extraction. In order to reduce consumption of model computation and memory, the channel pruning method was used to slim down the network. In order to improve the accuracy reduction caused by model pruning, the knowledge distillation of the teacher network-teaching assistant network-student network was adopted to the pruned network, and the student network can improve the recognition accuracy of alien invasive plants through learning the soft knowledge. In this study, the ablation experiments of the model were done. Three indicators including average accuracy, average recall rate and average F1 value were used to evaluate the classical models and the improved model MobileNet-LW. The results of ablation experiments showed that the performance of each improved method on the model was improved on a same testing seT. The accuracy of MobileNet-LW increased by 5.4 percentage points in identifying the alien invasive plants, and the number of parameters of model reduced by about 53%. The average accuracies of the five models, i.e., EfficentNet, DBTNet, ResNet-101, ConvNext and MobileNet-LW, were 72.3%, 74.9%, 76.1%, 79.7% and 86.1%, respectively, showing that the improved model could improve the identification accuracy of alien invasive plants. The alien invasive plant identification model based on the improved MobileNet showed high accuracy in identifying 113 species of alien invasive plants, and showed characteristic of lightweight after pruning. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
