The SF6 decomposed products are very alike,so the peaks of the SF6 decomposed products are overlapped.It make us hard to quantitatively calculate.This paper introduced a method to separate the overlapped chromatogram ...The SF6 decomposed products are very alike,so the peaks of the SF6 decomposed products are overlapped.It make us hard to quantitatively calculate.This paper introduced a method to separate the overlapped chromatogram peaks with iterative curve-fitting and Gauss function,and deduced the formula of the iterative curve-fitting;provided a method to calculate the initial value of the peaks intensity and width by iterative formula,and confirm the final peaks intensity and width to separate the overlapped signals.By the analysis of the real chromatogram data,it proved the validity of this method on the main components used in distinguishing the GIS internal defects.This method overcomes the influence of SF6 peak to the other thin concentration peaks,so it can be used in the study of the quantitative analysis of the decomposed products on different failures and different voltages.展开更多
针对分选线上樱桃叠果和表面瑕疵难以准确实时分割的问题,该研究提出了一种基于改进YOLOv8n-seg的多分支实时分割模型,命名SemIns-YOLOv8。该模型在YOLOv8n-seg的PAN-FPN(path aggregation network for feature pyramid network)模块后...针对分选线上樱桃叠果和表面瑕疵难以准确实时分割的问题,该研究提出了一种基于改进YOLOv8n-seg的多分支实时分割模型,命名SemIns-YOLOv8。该模型在YOLOv8n-seg的PAN-FPN(path aggregation network for feature pyramid network)模块后引入基于上下文集成的语义分割模块,并采用交叉熵损失与Dice Loss联合的损失函数,替代原始实例分割模块对果梗和瑕疵进行识别,既提升了尺寸较小及特征不明显瑕疵的识别精度,又缩短了图像的识别时间。同时,通过提高特征图分辨率并引入豪斯多夫距离损失(Hausdorff distance loss,HD Loss)构建边界特征增强的实例分割模块,实现了樱桃重叠果体的精准分离。试验结果表明,SemIns-YOLOv8在樱桃分割任务中果体mAP50-95(mean average precision at intersection over union thresholds from 0.50 to 0.95)、果梗IoU和瑕疵mIoU(mean intersection over union)分别为98.20%、92.15%和65.97%,与YOLOv8n-seg相比,提升了2.10、2.33和14.35个百分点,并且在模型输入尺寸为1024×384像素时,单帧推理时间为23 ms,可为线上水果外观品质实时分选提供参考。展开更多
基金Project Supported by National Natural Science Foundation of China （ 50777070）, Science and Technique Project of Chongqing （CSTC, 2007AC2041 ）.
文摘The SF6 decomposed products are very alike,so the peaks of the SF6 decomposed products are overlapped.It make us hard to quantitatively calculate.This paper introduced a method to separate the overlapped chromatogram peaks with iterative curve-fitting and Gauss function,and deduced the formula of the iterative curve-fitting;provided a method to calculate the initial value of the peaks intensity and width by iterative formula,and confirm the final peaks intensity and width to separate the overlapped signals.By the analysis of the real chromatogram data,it proved the validity of this method on the main components used in distinguishing the GIS internal defects.This method overcomes the influence of SF6 peak to the other thin concentration peaks,so it can be used in the study of the quantitative analysis of the decomposed products on different failures and different voltages.
文摘针对分选线上樱桃叠果和表面瑕疵难以准确实时分割的问题,该研究提出了一种基于改进YOLOv8n-seg的多分支实时分割模型,命名SemIns-YOLOv8。该模型在YOLOv8n-seg的PAN-FPN(path aggregation network for feature pyramid network)模块后引入基于上下文集成的语义分割模块,并采用交叉熵损失与Dice Loss联合的损失函数,替代原始实例分割模块对果梗和瑕疵进行识别,既提升了尺寸较小及特征不明显瑕疵的识别精度,又缩短了图像的识别时间。同时,通过提高特征图分辨率并引入豪斯多夫距离损失(Hausdorff distance loss,HD Loss)构建边界特征增强的实例分割模块,实现了樱桃重叠果体的精准分离。试验结果表明,SemIns-YOLOv8在樱桃分割任务中果体mAP50-95(mean average precision at intersection over union thresholds from 0.50 to 0.95)、果梗IoU和瑕疵mIoU(mean intersection over union)分别为98.20%、92.15%和65.97%,与YOLOv8n-seg相比,提升了2.10、2.33和14.35个百分点,并且在模型输入尺寸为1024×384像素时,单帧推理时间为23 ms,可为线上水果外观品质实时分选提供参考。
