|
| 基于双重相似度孪生网络的小样本实例分割 |
| Few-shot instance segmentation based on double similarity Siamese network |
| 投稿时间：2019-10-08 |
| DOI： |
| 中文关键词: 小样本学习 实例分割 孪生网络 残差网络 空域相似度 频域相似度 |
| 英文关键词: few-shot learning instance segmentation Siamese network residual network spatial similarity frequency similarity |
| 基金项目:国家自然科学基金资助项目(61773297). |
|
| 摘要点击次数: 4321 |
| 全文下载次数: 3058 |
| 中文摘要: |
| 针对孪生网络在小样本数据集上的应用和优化问题,提出一种基于双重相似度计算和孪生网络的小样本实例分割模型。首先对传统孪生网络进行改进,将孪生网络与残差网络相结合,构建作为本模型骨干网络的孪生残差网络；然后在相似度计算阶段构建了具有两个子网络的双重相似度计算网络,分别用于计算场景图像与参考图像的空域相似度和频域相似度,并进行相似度特征融合；最后通过实例分割网络获得分割结果。此外,还引入Focal Loss损失函数来解决模型训练过程中正、负样本以及难、易样本的不均衡问题。在COCO数据集上的实验结果表明,本文方法的小样本实例分割性能要优于对比算法。 |
| 英文摘要: |
| For the optimization of Siamese network applied to datasets with small samples, a few-shot instance segmentation model named as DSSN is proposed based on double similarity calculation and Siamese network. Firstly, traditional Siamese network and residual network are combined to build the Siamese residual network as the backbone network of DSSN. Secondly, a double similarity calculation network with two sub-networks is constructed, which is used to calculate the spatial and frequency similarities between a scene image and a reference one respectively, and then the two kinds of similarity features are fused. Finally, the segmentation result is obtained from the instance segmentation network. In addition, Focal Loss function is introduced to solve the problems of unbalanced positive and negative samples as well as unbalanced difficult and simple samples during the model training. The experimental results on COCO dataset show that the few-shot instance segmentation performance of DSSN is superior to that of the other two algorithms. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
