高级检索

基于深度半监督学习的植物叶片自动识别

Automatic Leaf Recognition Based on Deep Semi-supervised Learning

  • 摘要: 植物叶片自动识别算法在植物教学和生态保护等领域有着广泛应用, 但由于植物种类繁多且类间差异小, 传统深度学习方法需要大量数据标注才能获得较好的训练效果. 为此, 提出一种基于深度半监督学习的植物叶片自动识别方法. 首先基于一致性正则化思想, 为提升数据扰动质量设计了显性、隐性数据扰动流程; 然后运用深度特征提取网络DenseNet, 有效地提升植物叶片细粒度特征的辨别能力; 最后基于模拟退火训练策略过滤训练过程中的异常数据, 从而缓解过拟合现象. 在分别含有5 284幅和18 000幅植物叶片图像的公开数据集MalayaKew-D3和私有数据集LeafSZU-2021中获得的实验结果表明, 与全量标注数据下监督学习模型相比, 该方法在仅使用30%~50%标注数据量时, 仍能达到92.36%~96.85%植物叶片识别准确率. 在相同数据标注量下, 该方法比当前最新的半监督球面均值聚类方法的平均识别准确度提高了2.95%, 且模型参数量降低了38.12%, 识别速度提高了61.51%.

     

    Abstract: Automatic leaf recognition algorithms have been widely used in the fields of plant teaching and ecological protection. However, due to the wide variety of plants and the small differences between classes, traditional deep learning methods require a large amount of data annotations to obtain better training results. Therefore, a semi-supervised deep learning method for plant leaf automatic identification is proposed. First, based on the idea of consistent regularization, both of explicit and implicit data perturbation processes are designed to improve the quality of data perturbation. Secondly, by utilizing a deep feature extraction network DenseNet, the ability of fine detail recognition can be effectively improved. Finally, based on the simulated annealing training strategy, the abnormal data in the training process is filtered to alleviate the over-fitting phenomenon. Extensive experiments are conducted on a public dataset MalayaKew-D3 with 5 284 leaf images and a private dataset LeafSZU-2021 with 18 000 leaf images. Compared with the fully supervised learning methods, the proposed method can still achieve 92.36% to 96.85% accuracy of plant leaf recognition by using only 30% to 50% labeled data. Compared with the latest semi-supervised spherical mean clustering model under the same amount of data annotations, the average recognition accuracy of the proposed method improves by 2.95%, where the number of model parameters also decreases by 38.12%, and the recognition speed improves by 61.51%.

     

/

返回文章
返回