| 引用本文: | 吴健辉,何灿,何俊康,谢永芳,赵林,张国云.FSNet: 基于频率特性的烟雾图像分割网络[J].控制理论与应用,2023,40(4):702~712.[点击复制] |
| WU Jian-hui,HE Can,HE Jun-kang,XIE Yong-fang,ZHAO Lin,ZHANG Guo-yun.FSNet: A network for smoke image segmentation based on frequency characteristics[J].Control Theory and Technology,2023,40(4):702~712.[点击复制] |
|
| FSNet: 基于频率特性的烟雾图像分割网络 |
| FSNet: A network for smoke image segmentation based on frequency characteristics |
| 摘要点击 1572 全文点击 472 投稿时间：2021-09-10 修订日期：2022-05-17 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2022.10857 |
| 2023,40(4):702-712 |
| 中文关键词 烟雾图像分割 频率特性 空洞空间金字塔池化 多任务学习 深度学习 |
| 英文关键词 smoke image segmentation frequency characteristics atrous spatial pyramid pooling multi-task learning deep learning |
| 基金项目 湖南省自然科学基金项目(2019JJ40110, 2019JJ40104), 湖南省教育厅科研基金项目(18B349, 19A201, 20A223) |
|
| 中文摘要 |
| 烟雾图像分割是对烟雾进行识别与精准定位的基础, 是火灾预警的重要手段. 针对烟雾分割时存在过分割、欠分割以及边界拟合粗糙的问题, 本文提出一种基于频率分离特性的烟雾图像分割网络. 所提出的频率分离模块将特征图中的烟雾区域分离为低频主体部分和高频边界部分, 同时基于多任务学习设计多模块权重自适应损失函数对烟雾整体、主体、边界分别监督学习, 起到细化烟雾边界和改善烟雾整体分割结果的作用; 此外, 结合可变形卷积提出改进的空洞空间金字塔池化模块以解决其信息利用率低和特征关联性差的问题. 在对比实验中, FSNet的烟雾交并比为76.55%, 比基线网络提高了4.25%. 可视化分割结果可以看出, FSNet能有效缓解过分割、欠分割, 所得烟雾边界更平滑, 烟雾图像分割的整体性能获得较大提升. |
| 英文摘要 |
| Smoke image segmentation is the basis of smoke recognition and accurate localization, and is an important tool for fire warning. In existing algorithms for smoke image segmentation, there are some problems such as oversegmentation, under-segmentation, and rough boundary fitting. To mitigate these problems, a smoke image segmentation network based on the frequency separation characterization called as frequency separation network (FSNet) is proposed in this paper. The proposed frequency separation module divides the feature map of smoke region into a main body of low-frequency and a boundary of high-frequency. Also, a multi-module loss function with daptive weights based on the multi-task learning is designed to learn the features of main body, boundary and whole smoke region under the supervision of the loss, refining the smoke boundary and improving the overall segmentation results of the smoke image. In addition, a modified atrous spatial pyramid pooling module is proposed to solve the problems of low information utilization and poor feature correlation by combining the deformable convolution. In the comparison experiments, the intersection over union (IoU) of FSNet is 76.55%, which is 4.25% higher than the baseline network. The visualization results show that the FSNet can effectively mitigate over-segmentation and under-segmentation, which helps to obtain smoother smoke boundary and significantly improves the overall performance of smoke image segmentation. |
|
|
|
|
|