| 引用本文: | 张大锦,刘辉,陈甫刚,赵安.频域多方向C-UNet及动态损失的工业烟尘图像分割[J].控制理论与应用,2024,41(3):543~554.[点击复制] |
| ZHANG Da-jin,LIU Hui,CHEN Fu-gang,ZHAO An.Industrial smoke image segmentation based on frequency domain multi-directional C-UNet and dynamic loss[J].Control Theory and Technology,2024,41(3):543~554.[点击复制] |
|
| 频域多方向C-UNet及动态损失的工业烟尘图像分割 |
| Industrial smoke image segmentation based on frequency domain multi-directional C-UNet and dynamic loss |
| 摘要点击 2340 全文点击 61 投稿时间：2022-11-07 修订日期：2024-01-15 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2023.20987 |
| 2024,41(3):543-554 |
| 中文关键词 工业烟尘 图像分割 轮廓波变换 特征提取 动态损失函数 |
| 英文关键词 industrial smoke image segmentation contourlet transform feature extraction dynamic loss function |
| 基金项目 国家自然科学基金项目(62263016, 61863018), 云南省科技厅应用基础研究项目(202001AT070038)资助. |
|
| 中文摘要 |
| 工业烟尘污染等级监测中烟尘的准确分割是污染等级判定的重要前提. 针对边缘模糊且方向多变烟尘在
特征提取过程中边缘方向细节信息提取困难、分割不准确的问题, 本文提出一种频域多方向C-UNet及动态损失的
工业烟尘图像分割方法. 首先, 通过构建轮廓波多方向分解下采样结构增强编码阶段烟尘边缘方向信息的提取能
力; 其次, 通过轮廓波变换提取烟尘8个边缘方向细节信息进行跳跃连接, 提升持续采样过程中细节信息的表达准
确度; 然后, 构建轮廓波细节重构上采样结构增强解码阶段烟尘边缘细节信息的恢复能力; 最后, 提出一种动态加
权策略构建组合损失函数来优化训练网络, 增强网络对烟尘边缘特征的提取能力. 结果表明, 本文方法与U-Net和
其他同类方法相比在指标上有较好提升, 改善了烟尘边缘分割不准确的问题, 在不同烟尘场景上的分割效果也优于
现有分割模型. |
| 英文摘要 |
| The accurate segmentation of the smoke in industrial smoke pollution level monitoring is an important prerequisite
for pollution level determination. The typical challenges in feature extraction of smoke include blurred edges,
difficult extraction of edge directional detail information and inaccurate segmentation. In this study, a frequency domain
multi-directional C-UNet (Contourlet U-Net) and dynamic loss industrial smoke image segmentation method is proposed,
aiming to provide support to overcome these problems. Firstly, the contourlet multi-directional decomposition downsampling
structure is constructed to enhance the ability to extract edge direction information of smoke in the encoding
stage. Secondly, the contourlet transform is used to extract detailed information on the eight edge directions of smoke for
skip connections, improving the accuracy of detail information expression during continuous sampling. Then, the contourlet
detail reconstruction up-sampling structure is constructed to enhance the recovery ability of edge detail information
of smoke in the decoding stage. Finally, a dynamic weighting strategy is proposed to construct a combined loss function
to optimize the training network and enhance the network’s ability to extract smoke edge features. The results show that
compared with U-Net and other similar methods, the proposed method has a better improvement in indicators, improves
the accuracy of smoke edge segmentation, and the segmentation effect on different smoke scenes is better than the existing
segmentation model. |
|
|
|
|
|