The detection of individual pigs and their parts is a key step to realizing automatic recognition of group-housed pigs’behavior by video monitoring.However,it is still difficult to accurately locate each individual p...The detection of individual pigs and their parts is a key step to realizing automatic recognition of group-housed pigs’behavior by video monitoring.However,it is still difficult to accurately locate each individual pig and their body parts from video images of groups-housed pigs.To solve this problem,a Cascade Faster R-CNN Pig Detector(C-FRPD)was designed to detect the individual pigs and different parts of their body.Firstly,the features were extracted by 101-layers Residual Networks(ResNet-101)from video images of group-housed pigs,and the features were input into the region proposal networks(RPN)to obtain the region proposals.Then classification and bounding box regression on region proposals were performed to get the location of each pig.Finally,the body parts of the pig were determined by using the Cascade structure to search on the feature map of the pig body area.These operations completed the detection of the whole body of each pig and its different parts of the body,and established the association between the whole and parts of body in the end-to-end detection.In this study,1500 pig pen images were trained and tested.The test results showed that the detection accuracy of C-FRPD reached 98.4%.Compared with the Faster R-CNN without cascade structure,the average detection accuracy was increased by 4.3 percentage points.The average detection time of a single image was 259 ms.The method in this study could accurately detect and correlate the individual pig with its head,back,and tail in the image.This method can provide a technical reference for recognizing the behavior of group-housed pigs.展开更多
针对光照不均、噪声大、拍摄质量不高的夜晚水下环境,为实现夜晚水下图像中鱼类目标的快速检测,利用计算机视觉技术,提出了一种基于改进Cascade R-CNN算法和具有色彩保护的MSRCP(Multi-scale Retinex with color restoration)图像增强...针对光照不均、噪声大、拍摄质量不高的夜晚水下环境,为实现夜晚水下图像中鱼类目标的快速检测,利用计算机视觉技术,提出了一种基于改进Cascade R-CNN算法和具有色彩保护的MSRCP(Multi-scale Retinex with color restoration)图像增强算法的夜晚水下鱼类目标检测方法。首先针对夜晚水下环境的视频数据,根据时间间隔,截取出相应的夜晚水下鱼类图像,对截取的原始图像进行MSRCP图像增强。然后采用DetNASNet主干网络进行网络训练和水下鱼类特征信息的提取,将提取出的特征信息输入到Cascade R-CNN模型中,并使用Soft-NMS候选框优化算法对其中的RPN网络进行优化,最后对夜晚水下鱼类目标进行检测。实验结果表明,该方法解决了夜晚水下环境中的图像降质、鱼类目标重叠检测问题,实现了对夜晚水下鱼类目标的快速检测,对夜晚水下鱼类图像目标检测的查准率达到95.81%,比Cascade R-CNN方法提高了11.57个百分点。展开更多
基金This work was financially supported by the Key Research and Development Program of Guangdong Province(Grant No.2019B0202150042019B090922002)the Agricultural Research Project and Agricultural Technology Promotion Project of Guangdong(Grant No.2021KJ383)。
文摘The detection of individual pigs and their parts is a key step to realizing automatic recognition of group-housed pigs’behavior by video monitoring.However,it is still difficult to accurately locate each individual pig and their body parts from video images of groups-housed pigs.To solve this problem,a Cascade Faster R-CNN Pig Detector(C-FRPD)was designed to detect the individual pigs and different parts of their body.Firstly,the features were extracted by 101-layers Residual Networks(ResNet-101)from video images of group-housed pigs,and the features were input into the region proposal networks(RPN)to obtain the region proposals.Then classification and bounding box regression on region proposals were performed to get the location of each pig.Finally,the body parts of the pig were determined by using the Cascade structure to search on the feature map of the pig body area.These operations completed the detection of the whole body of each pig and its different parts of the body,and established the association between the whole and parts of body in the end-to-end detection.In this study,1500 pig pen images were trained and tested.The test results showed that the detection accuracy of C-FRPD reached 98.4%.Compared with the Faster R-CNN without cascade structure,the average detection accuracy was increased by 4.3 percentage points.The average detection time of a single image was 259 ms.The method in this study could accurately detect and correlate the individual pig with its head,back,and tail in the image.This method can provide a technical reference for recognizing the behavior of group-housed pigs.
文摘热轧带钢是钢铁行业的重要产品,其表面缺陷是影响产品质量的重要因素。针对传统缺陷检测算法存在的过程繁琐、精度不足和效率低下等问题,提出一种基于改进更快速区域卷积神经网络(faster region-based convolutional neural network,Faster R-CNN)的检测算法,实现对热轧带钢表面缺陷的高效、高精度检测。首先,采用特征相加的方法对底层细节特征和高层语义特征进行融合;然后,采用精准的感兴趣区域池化(precise region of interest pooling,Precise ROI Pooling)获取固定大小的特征向量,避免特征出现位置偏差;最后,利用均值偏移聚类算法对带钢数据集进行聚类,获得适用于热轧带钢表面缺陷检测的先验框尺寸。实验结果表明,所提算法在热轧带钢表面缺陷检测数据集上的平均精度均值达到了85.34%,检测速度为23.5帧/s,且鲁棒性良好,满足实际的工业检测需求。
文摘针对光照不均、噪声大、拍摄质量不高的夜晚水下环境,为实现夜晚水下图像中鱼类目标的快速检测,利用计算机视觉技术,提出了一种基于改进Cascade R-CNN算法和具有色彩保护的MSRCP(Multi-scale Retinex with color restoration)图像增强算法的夜晚水下鱼类目标检测方法。首先针对夜晚水下环境的视频数据,根据时间间隔,截取出相应的夜晚水下鱼类图像,对截取的原始图像进行MSRCP图像增强。然后采用DetNASNet主干网络进行网络训练和水下鱼类特征信息的提取,将提取出的特征信息输入到Cascade R-CNN模型中,并使用Soft-NMS候选框优化算法对其中的RPN网络进行优化,最后对夜晚水下鱼类目标进行检测。实验结果表明,该方法解决了夜晚水下环境中的图像降质、鱼类目标重叠检测问题,实现了对夜晚水下鱼类目标的快速检测,对夜晚水下鱼类图像目标检测的查准率达到95.81%,比Cascade R-CNN方法提高了11.57个百分点。
