一种基于彩票迁移的稀疏网络植株病虫害识别模型

A lottery ticket-based deep transfer sparse neural network for plant disease identification

植株病虫害的识别能够有效地提高农作物产量,当前数据驱动的深度植株病虫害识别方法需要大量的有标签数据,导致现有方法难以很好地识别少样本的新病虫。且基于深度学习的方法需要对大量的参数进行训练,难以削减计算开销。研究设计一种基于彩票迁移的稀疏网络植株病虫害识别模型:定义深度网络的彩票迁移假设,利用压缩策略构建稀疏网络,识别迁移源域的本质知识,提高深度网络的迁移效率;然后,设计深度彩票迁移算法,训练植株病虫害深度识别模型,解决少样本病虫识别调整;最后,在典型的通用数据与植株病虫害识别数据集上,验证基于彩票迁移的深度植株病虫害识别模型能高效迁移源域的本质知识。在PlantVillage数据集上,对植株病虫害识别准确率为97.69%,且所需训练的参数只有原始网络的约30%。

In agriculture,the plant disease identification can increase the production of crops.The existing data-driven deep plant disease identification methods are based on a great number of supervised data,posing vast challenges on detecting new pests of few data.And there are many trainable parameters in those deep learning-based methods,costing much computation resources.To solve those challenges,a lottery ticket-based deep sparse transfer method is proposed for the plant disease identification.Specifically,the deep lottery ticket hypothesis is introduced,in which a compressing strategy is designed to construct the deep sparse network that distills useful information in the auxiliary domains,improving the transfer efficiency.Then,a deep lottery ticket transfer algorithm is proposed to train a deep plant disease identification model that can effectively detect the new pests of few data.Finally,the proposed method is evaluated on the representative datasets,i.e.,CIFAR-10 and PlantVillage,and the accuracy of detecting new pests can achieve 97.69%in plantViuage with 70%-parameter-reduction.