郝志航, 张小咏, 陈正超,等.基于改进的Faster R-CNN污水处理厂目标提取[J].河南理工大学学报（自然科学版）,2024,43(1):68-77.

HAO Z H, ZHANG X Y, CHEN Z C, et al.Target extraction of sewage treatment plant based on improved Faster R-CNN[J].Journal of Henan Polytechnic University(Natural Science) ,2024,43(1):68-77.

基于改进的Faster R-CNN污水处理厂目标提取

郝志航1, 张小咏1, 陈正超2, 卢凯旋2

1.北京信息科技大学 高动态导航技术北京市重点实验室，北京  100101; 2.中国科学院 空天信息创新研究院，北京  100094

摘要:目的 为解决传统污水处理厂检测耗时耗力，很难满足大范围、高频次污水处理厂监测的问题，方法 选取国产高分二号卫星影像（GF-2）数据作为样本制作来源，选择京津冀地区作为研究区域，以深度学习技术作为基础，提出一种面向污水处理厂目标提取的自适应可变形卷积网络（adaptive deformable convolution network，ADCN）。结果 消融实验结果表明，随着卷积神经网络深度逐步递增，模型精确度和召回率均有所提高；通过特征金字塔融合的多尺度特征，有效弥补了小目标漏检的缺陷，ADCN在以上基础上增加的可变形卷积和可变形区域池化，在提高精度的同时，可明显改善边框的回归精度，ADCN在精度为85%的前提下，召回率达到95.1%；对比实验表明，相比于SSD，YOLO，Retinanet，Faster R-CNN算法，ADCN的mAP精度最高，达95.32%，同时该算法在大、中、小三种尺度的污水处理厂提取的结果中表现优异；通过ADCN对京津冀地区的污水处理厂进行提取，共提取京津冀地区污水处理厂152个，其中北京15个、天津26个，河北111个，人工对比后误检为17个，检出率为92.68%。结论通过结合深度学习技术和遥感影像数据，可以大范围、快速提取污水处理厂目标，有效解决传统污水处理厂检测耗时问题，提高对污水处理厂的管理和监控。

关键词:深度学习;目标检测;污水处理厂目标提取;京津冀地区;可变形卷积

doi:doi：10.16186/j.cnki.1673-9787.2021100063

基金项目:国家自然科学基金资助项目（41871348）；国家科技重大专项项目（03-Y30F03-9001-20/22）

收稿日期:2021/10/26

修回日期:2022/02/17

出版日期:2024/01/25

Target extraction of sewage treatment plant based on improved Faster R-CNN

HAO Zhihang1, ZHANG Xiaoyong1, CHEN Zhengchao2, LU Kaixuan2

1.Beijing Key Laboratory of High Dynamic Navigation Technology，Beijing University of Information Technology，Beijing  100101，China; 2.Aerospace Information Research Institute，Chinese Academy of Sciences，Beijing  100094，China

Abstract: Objective There is a problem of time-consuming and labor-intensive testing in traditional sewage treatment plants，which makes it difficult to meet the needs of large-scale and high-frequency monitoring of sewage treatment plants. Methods Using domestic GF-2 satellite imagery data as the sample production source，the Beijing-Tianjin-Hebei Region was selected as the research area. Based on deep learning technology，a self-adaptive deformable convolutional network（adaptive deformable convolution network，ADCN） for target extraction of sewage treatment plants was proposed. Results The ablation experiment results show that as the depth of the convolutional neural network gradually increases，the accuracy and recall rate of the model are both improved. The multi-scale features fused through the feature pyramid effectively compensate for the defect of small target missed detection. The deformable convolution and deformable region pooling added by ADCN on the basis of the above，which can significantly improve the regression accuracy of the bounding box while improving the accuracy. ADCN can achieve a recall rate of 95.1% with an accuracy of 85%. Comparative experiments have shown that compared to SSD，YOLO，Retinanet，Faster R-CNN algorithms，the ADCN network has the best accuracy on mAP，reaching 95.32%. Excellent performance was observed in the extraction results from sewage treatment plants at three scales：large，medium，and small. Finally，152 sewage treatment plants in the Beijing-Tianjin-Hebei Region were extracted through the ADCN network，including 15 in Beijing，26 in Tianjin，and 111 in Hebei. After manual comparison，there were 17 faise detection，with a detection rate of 92.68%. Conclusion By combining deep learning technology and remote sensing image data，it is possible to quickly extract targets from sewage treatment plants on a large scale，effectively solving the time-consuming problem of traditional sewage treatment plant detection，and improving the management and monitoring of sewage treatment plants.

Key words:deep learning;object detection;sewage treatment plant extraction;Beijing-Tianjin-Hebei Region;deformable convolution