基于多任务持续学习的树种识别

Identification of Tree Species Based on Multi-Task Continual Learning

针对现有的深度学习方法在树干或树叶单一识别任务上需要大量样本做标注和训练的问题,且存在灾难性遗忘现象,提出一种新的神经网络模型用于多任务树种识别。对于少量不同类型数据样本,本文引入持续学习,将树干识别和树叶识别看作2个连续的学习,实现多任务识别。训练模型分为2个阶段:第1阶段为树干识别,保留参数重要性;第2阶段引入正则化损失约束重要参数的变化,维持模型对于叶片的特征提取能力,而保持低重要性参数的改变,以学习不同树种样本中更多的特征信息。测试结果表明,该方法在树干识别和树叶识别时的准确率分别为91.75%和98.85%,较单任务的深度学习有18.03%和11.92%的提升。本研究所提出的模型更适用于在不同样本中进行多任务分类识别,较好地避免灾难性遗忘问题。

For the current problem that methods of deep learning need a large number of examples for labeling and training in solving single task such as trunk recognition or leaf recognition,and there is a phenomenon called catastrophic forgetting,this paper has proposed a new neural network model for multi-task tree species recognition.For a small number of different data samples,continual learning is introduced in this paper,and trunk recognition and leaf recognition are regarded as two continuous learning to accomplish multi-task recognition.The training model is divided into two stages:trunk recognition is settled in the first stage,which preserves the significance of parameters.In the second stage,regularization loss constraint is introduced to maintain the feature extraction ability of the model for leaves,while keeping the changes of low importance parameters,so as to learn more feature information from different tree species samples.The experiment shows that the accuracy in trunk recognition and leaf recognition are 91.75%and 98.85%,respectively,which are 18.03%and 11.92%higher than single-task deep learning.It indicates that the proposed model is more suitable for multi-task classification and recognition in different samples,and it effectively prevents catastrophic forgetting in deep learning.