针对复杂海洋环境中存在背景噪声、海洋垃圾特征模糊和目标尺度小的检测挑战,本文提出一种基于改进CenterNet的海洋垃圾无锚检测算法——MG-CenterNet。引入GB(green-blue)注意力机制,通过关注海洋图像绿色、蓝色通道来增强特征提取;利...针对复杂海洋环境中存在背景噪声、海洋垃圾特征模糊和目标尺度小的检测挑战,本文提出一种基于改进CenterNet的海洋垃圾无锚检测算法——MG-CenterNet。引入GB(green-blue)注意力机制,通过关注海洋图像绿色、蓝色通道来增强特征提取;利用跨层特征聚合(cross-layer feature aggregation,CFA)模块丰富关键特征反馈,使模型获取更多像素级语义信息从而精准分类图像;构造完全交并比(complete intersection over union,CIoU)损失函数优化边界框匹配度,进一步提高目标定位精度。MG-CenterNet在TrashCan数据集和自建数据集上分别取得了77.98%和76.92%的平均精确率均值(mean average precision,m AP),推理速度分别达到27.18帧/s和26.98帧/s。研究结果证明MG-CenterNet在检测精度上显著优于其他算法,满足实时检测的要求。低对比度及遮挡条件下的验证实验进一步证明了所提出算法的鲁棒性和可靠性,为复杂环境中的海洋垃圾检测提供了科学参考。展开更多
针对小尺度目标在检测时精确率低且易出现漏检和误检等问题,提出一种改进的YOLOv3(You Only Look Once version 3)小目标检测算法。在网络结构方面,为提高基础网络的特征提取能力,使用DenseNet-121密集连接网络替换原Darknet-53网络作...针对小尺度目标在检测时精确率低且易出现漏检和误检等问题,提出一种改进的YOLOv3(You Only Look Once version 3)小目标检测算法。在网络结构方面,为提高基础网络的特征提取能力,使用DenseNet-121密集连接网络替换原Darknet-53网络作为其基础网络,同时修改卷积核尺寸,进一步降低特征图信息的损耗,并且为增强检测模型对小尺度目标的鲁棒性,额外增加第4个尺寸为104×104像素的特征检测层;在对特征图融合操作方面,使用双线性插值法进行上采样操作代替原最近邻插值法上采样操作,解决大部分检测算法中存在的特征严重损失问题;在损失函数方面,使用广义交并比(GIoU)代替交并比(IoU)来计算边界框的损失值,同时引入Focal Loss焦点损失函数作为边界框的置信度损失函数。实验结果表明,改进算法在VisDrone2019数据集上的均值平均精度(mAP)为63.3%,较原始YOLOv3检测模型提高了13.2百分点,并且在GTX 1080 Ti设备上可实现52帧/s的检测速度,对小目标有着较好的检测性能。展开更多
在自然场景中,天气情况、光照强度、背景干扰等问题影响火焰检测的准确性.为了实现复杂场景下实时准确的火焰检测,在目标检测网络YOLOv5的基础上,结合Focal Loss焦点损失函数、CIoU(Complete Intersection over Union)损失函数与多特征...在自然场景中,天气情况、光照强度、背景干扰等问题影响火焰检测的准确性.为了实现复杂场景下实时准确的火焰检测,在目标检测网络YOLOv5的基础上,结合Focal Loss焦点损失函数、CIoU(Complete Intersection over Union)损失函数与多特征融合,提出实时高效的火焰检测方法.为了缓解正负样本不均衡问题,并充分利用困难样本的信息,引入焦点损失函数,同时结合火焰静态特征和动态特征,设计多特征融合方法,达到剔除误报火焰的目的.针对国内外缺乏火焰数据集的问题,构建大规模、高质量的十万量级火焰数据集(http://www.yongxu.org/data bases.html).实验表明,文中方法在准确率、速度、精度和泛化能力等方面均有明显提升,同时降低误报率.展开更多
The contact network dropper works in a harsh environment,and suffers from the impact effect of pantographs during running of trains,which may lead to faults such as slack and broken of the dropper wire and broken of t...The contact network dropper works in a harsh environment,and suffers from the impact effect of pantographs during running of trains,which may lead to faults such as slack and broken of the dropper wire and broken of the current-carrying ring.Due to the low intelligence and poor accuracy of the dropper fault detection network,an improved fully convolutional one-stage(FCOS)object detection network was proposed to improve the detection capability of the dropper condition.Firstly,by adjusting the parameterαin the network focus loss function,the problem of positive and negative sample imbalance in the network training process was eliminated.Secondly,the generalized intersection over union(GIoU)calculation was introduced to enhance the network’s ability to recognize the relative spatial positions of the prediction box and the bounding box during the regression calculation.Finally,the improved network was used to detect the status of dropper pictures.The detection speed was 150 sheets per millisecond,and the MAP of different status detection was 0.9512.Through the simulation comparison with other object detection networks,it was proved that the improved FCOS network had advantages in both detection time and accuracy,and could identify the state of dropper accurately.展开更多
文摘针对复杂海洋环境中存在背景噪声、海洋垃圾特征模糊和目标尺度小的检测挑战,本文提出一种基于改进CenterNet的海洋垃圾无锚检测算法——MG-CenterNet。引入GB(green-blue)注意力机制,通过关注海洋图像绿色、蓝色通道来增强特征提取;利用跨层特征聚合(cross-layer feature aggregation,CFA)模块丰富关键特征反馈,使模型获取更多像素级语义信息从而精准分类图像;构造完全交并比(complete intersection over union,CIoU)损失函数优化边界框匹配度,进一步提高目标定位精度。MG-CenterNet在TrashCan数据集和自建数据集上分别取得了77.98%和76.92%的平均精确率均值(mean average precision,m AP),推理速度分别达到27.18帧/s和26.98帧/s。研究结果证明MG-CenterNet在检测精度上显著优于其他算法,满足实时检测的要求。低对比度及遮挡条件下的验证实验进一步证明了所提出算法的鲁棒性和可靠性,为复杂环境中的海洋垃圾检测提供了科学参考。
文摘针对小尺度目标在检测时精确率低且易出现漏检和误检等问题,提出一种改进的YOLOv3(You Only Look Once version 3)小目标检测算法。在网络结构方面,为提高基础网络的特征提取能力,使用DenseNet-121密集连接网络替换原Darknet-53网络作为其基础网络,同时修改卷积核尺寸,进一步降低特征图信息的损耗,并且为增强检测模型对小尺度目标的鲁棒性,额外增加第4个尺寸为104×104像素的特征检测层;在对特征图融合操作方面,使用双线性插值法进行上采样操作代替原最近邻插值法上采样操作,解决大部分检测算法中存在的特征严重损失问题;在损失函数方面,使用广义交并比(GIoU)代替交并比(IoU)来计算边界框的损失值,同时引入Focal Loss焦点损失函数作为边界框的置信度损失函数。实验结果表明,改进算法在VisDrone2019数据集上的均值平均精度(mAP)为63.3%,较原始YOLOv3检测模型提高了13.2百分点,并且在GTX 1080 Ti设备上可实现52帧/s的检测速度,对小目标有着较好的检测性能。
基金supported by Natural Science Foundation of Gansu Province(No.20JR10RA216)。
文摘The contact network dropper works in a harsh environment,and suffers from the impact effect of pantographs during running of trains,which may lead to faults such as slack and broken of the dropper wire and broken of the current-carrying ring.Due to the low intelligence and poor accuracy of the dropper fault detection network,an improved fully convolutional one-stage(FCOS)object detection network was proposed to improve the detection capability of the dropper condition.Firstly,by adjusting the parameterαin the network focus loss function,the problem of positive and negative sample imbalance in the network training process was eliminated.Secondly,the generalized intersection over union(GIoU)calculation was introduced to enhance the network’s ability to recognize the relative spatial positions of the prediction box and the bounding box during the regression calculation.Finally,the improved network was used to detect the status of dropper pictures.The detection speed was 150 sheets per millisecond,and the MAP of different status detection was 0.9512.Through the simulation comparison with other object detection networks,it was proved that the improved FCOS network had advantages in both detection time and accuracy,and could identify the state of dropper accurately.
