Maize residue segmentation using Siamese domain transfer network

文献类型： 外文期刊

第一作者： Li, Lin

作者： Li, Lin;Li, Jia;Lv, Chengxu;Yuan, Yanwei;Zhao, Bo

作者机构：

关键词： Maize residue; Few-shot samples; Intermediate domain; Domain distance; Transfer

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：5.565； 五年影响因子：5.494 ）

ISSN： 0168-1699

年卷期： 2021 年 187 卷

页码：

收录情况： SCI

摘要： Maize residue detection is an essential factor for conservation tillage; however, traditional methods are not robust for complex environments. Deep neural networks can learn multi-scale features from large datasets; nevertheless, it is difficult to obtain sufficient labelled maize residue data. Hence, it is important to develop an approach that can realize maize residue segmentation from few-shot samples. In this paper, a Siamese domain transfer network (SDTN) architecture is proposed to transfer convolution features to maize residue segmentation. First, an intermediate domain is used to bridge the distance between the source and target domains, which is a typical limiting factor in domain transfer. Next, the knowledge is transferred via the SDTN, in which hidden convolution layers represent multi-scale feature maps. Specific layers are explicitly matched using different domain distributions with designated confusion distances. Finally, the network is annealed and trained with fewshot samples. The experiments indicate that our transfer strategy and architecture can effectively transfer features and achieve a segmentation error of 33.2 AP at 5 fps. Furthermore, our model can reduce the error obtained by direct fine-tuning without domain transfer by 4.6%.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序