河道排口是水体污染的直接源头,无人机、无人船等巡检影像中的排口检测需人工解译,存在效率低、精度不足等问题。为实现智能识别,本文对公开发布的河道排口影像数据集iSOOD(Images for Sewage Outfalls Objective Detection)进行重新分...河道排口是水体污染的直接源头,无人机、无人船等巡检影像中的排口检测需人工解译,存在效率低、精度不足等问题。为实现智能识别,本文对公开发布的河道排口影像数据集iSOOD(Images for Sewage Outfalls Objective Detection)进行重新分类和扩充,构建了包含12658个排口样本的新数据集,并基于该数据集对比分析了Faster R-CNN(Faster Region-based Convolutional Neural Network)、SSDLite(Single Shot MultiBox Detector Lite)、RTMDet-tiny(Real-time Tiny Models for Object Detection)、YOLOv5n(You Only Look Once version 5 nano)和YOLOv8n(You Only Look Once version 8 nano)5种深度学习模型的排口检测效果,为进一步提升模型性能,以YOLOv5n模型为基础,通过引入不同注意力机制对其进行了对比试验。试验结果表明,YOLOv5n模型的平均精度均值(mean Average Precision,mAP)可达87.80%,高于其他模型0.4%~11.7%,不同注意力机制对YOLOv5n模型平均精度均值的提升在-1.13%~1.52%,且引入高效多尺度注意力(Efficient Multi-Scale Attention,EMA)注意力机制后,可改善部分遮挡及小目标排口的检测效果,将模型mAP提升至89.32%。展开更多
G-quadruplexes （G4s） play important roles in biological systems, such as telomere maintenance, replication, and transcription. Based on the DNA sequence, loop geometry, and the local environments, G4s can be classif...G-quadruplexes （G4s） play important roles in biological systems, such as telomere maintenance, replication, and transcription. Based on the DNA sequence, loop geometry, and the local environments, G4s can be classified into different conformations. It is important to detect different types of G4s to monitor the diseases related with G4s. Most ligands bind to G4s based on end-stacking modes, while rare ligands bind to G4s through groove binding modes. We have found that a cyanine dye DMSB interacts with parallel G4 by end-stacking and groove simultaneous binding mode. In this article, we found that DMSB could simply discriminate parallel G4s from other DNA motifs by using UV-vis spectrum. These results give some clues to develop high specificity G4 probes.展开更多
文摘河道排口是水体污染的直接源头,无人机、无人船等巡检影像中的排口检测需人工解译,存在效率低、精度不足等问题。为实现智能识别,本文对公开发布的河道排口影像数据集iSOOD(Images for Sewage Outfalls Objective Detection)进行重新分类和扩充,构建了包含12658个排口样本的新数据集,并基于该数据集对比分析了Faster R-CNN(Faster Region-based Convolutional Neural Network)、SSDLite(Single Shot MultiBox Detector Lite)、RTMDet-tiny(Real-time Tiny Models for Object Detection)、YOLOv5n(You Only Look Once version 5 nano)和YOLOv8n(You Only Look Once version 8 nano)5种深度学习模型的排口检测效果,为进一步提升模型性能,以YOLOv5n模型为基础,通过引入不同注意力机制对其进行了对比试验。试验结果表明,YOLOv5n模型的平均精度均值(mean Average Precision,mAP)可达87.80%,高于其他模型0.4%~11.7%,不同注意力机制对YOLOv5n模型平均精度均值的提升在-1.13%~1.52%,且引入高效多尺度注意力(Efficient Multi-Scale Attention,EMA)注意力机制后,可改善部分遮挡及小目标排口的检测效果,将模型mAP提升至89.32%。
基金supported by Major National Basic Research Projects (973,No.2013CB733701)National Natural Science Foundation of China (Nos.81072576,91027033,21302188,21205121,21305145 and 31200576)Chinese Academy of Sciences (No.KJCX2-EW-N06-01)
文摘G-quadruplexes （G4s） play important roles in biological systems, such as telomere maintenance, replication, and transcription. Based on the DNA sequence, loop geometry, and the local environments, G4s can be classified into different conformations. It is important to detect different types of G4s to monitor the diseases related with G4s. Most ligands bind to G4s based on end-stacking modes, while rare ligands bind to G4s through groove binding modes. We have found that a cyanine dye DMSB interacts with parallel G4 by end-stacking and groove simultaneous binding mode. In this article, we found that DMSB could simply discriminate parallel G4s from other DNA motifs by using UV-vis spectrum. These results give some clues to develop high specificity G4 probes.
