| 肖刚*,吴沛熙**,丁浩南**,陈锋***,徐雪松*,袁牧**,程振波*.基于域对抗的自适应环境运动目标状态检测方法[J].高技术通讯(中文),2026,36(1):67~79 |
| 基于域对抗的自适应环境运动目标状态检测方法 |
| Domain-adaptive environmental fish target status detection method |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 01. 006 |
| 中文关键词: 生物水质监测； 域适应； 目标检测； 前景融合 |
| 英文关键词: biological water quality monitoring, domain adaptation, target detection, foreground fusion |
| 基金项目: |
| 作者 | 单位 | | 肖刚* | （*浙江工业大学计算机科学与技术学院杭州 310023）
（**浙江工业大学机械工程学院杭州 310023）
（***浙江天行健水务有限公司杭州 310005） | | 吴沛熙** | | | 丁浩南** | | | 陈锋*** | | | 徐雪松* | | | 袁牧** | | | 程振波* | |
|
| 摘要点击次数: 22 |
| 全文下载次数: 15 |
| 中文摘要: |
| 生物式水质预警系统常将鱼类作为生物指示器，通过自动获取监测箱内鱼目标的行为状态，实现对流经监测箱水体的源水水质预警。然而，在源水杂质、藻类和水垢等长时间的影响下，使用现有的深度学习框架检测监测箱内的鱼目标，其准确性会逐渐下降。为此，本文提出了一种自适应环境鱼目标检测模型，该模型包括前景融合处理模块、域对抗模块以及目标检测模块。前景融合模块将输入图像与其包含的目标轮廓二值图像融合作为模型的输入，域对抗模块中的域分类网络经由梯度反转层来实现对输入图像所处域的分类，目标检测模块通过弱监督来增加不同域下训练数据的标注信息。最后，通过改变监测箱背景构建了5类不同水体环境的数据集并在这些数据集上进行实验，结果表明模型在监测箱水体环境发生变化的情况下，依然能较准确地对鱼目标进行状态分类。 |
| 英文摘要: |
| The biological water quality early warning systems commonly utilize fish as bio-indicators, implementing early warnings for the water quality of source water flowing through the monitoring chamber by automatically detecting the behavioral states of fish targets within the chamber. However, during prolonged operation, due to factors such as impurities, algae, and scale buildup in the source water, the accuracy of fish target detection within the monitoring chamber tends to decrease when using existing deep learning frameworks. To address this issue, this paper proposes an adaptive environmental fish target detection model, which includes a foreground fusion module, a domain adversarial module, and a target detection module. The foreground fusion module combines the input image with its corresponding binary image of the target contours as the model input. The domain adversarial module uses a domain classification network with a gradient reversal layer to classify the domain of the input image. The target detection module enhances the annotation information of training data in different domains through weak supervision. Finally, we construct five different aquatic environment datasets by altering the background of the monitoring chamber and conducted experiments on these datasets. Experimental results show that the model can accurately classify the status of fish targets even when the environmental conditions in the monitoring chamber change. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
