Pulp fiber length distribution characterization has been examined in this study. Because of the fiber morphology: slender in shape, fiber size distribution characterization is a very difficult task. Traditional techni...Pulp fiber length distribution characterization has been examined in this study. Because of the fiber morphology: slender in shape, fiber size distribution characterization is a very difficult task. Traditional technique involves separation of the particles by size, such as Bauer-McNett fiber classifier, and measuring the weight fractions. The particle fractions obtained may or may not reflect the desired size classification. On the other hand, the more recent technique through optical measurement of fiber length is limited by its inability to measure the mass of the particle fractions. Therefore, not only the two techniques fail to generate identical results, either one was accepted to be of better value. Pure hardwood kraft, softwood kraft, and their mixture samples have been measured for their fiber length distributions using an optical fiber quality analyzer: FQA. The data obtained from FQA are extensively studied to investigate more reliable way of representing the fiber length data and thus examining the viable route for measuring the fiber size distributions. It has been found that the fiber length averaged length l1 is a viable indicator of the average pulp fiber length. The fiber size fraction and/or distribution can be represented by the fiber "length" fractions.展开更多
目的图表问答是计算机视觉多模态学习的一项重要研究任务,传统关系网络(relation network,RN)模型简单的两两配对方法可以包含所有像素之间的关系,因此取得了不错的结果,但此方法不仅包含冗余信息,而且平方式增长的关系配对的特征数量...目的图表问答是计算机视觉多模态学习的一项重要研究任务,传统关系网络(relation network,RN)模型简单的两两配对方法可以包含所有像素之间的关系,因此取得了不错的结果,但此方法不仅包含冗余信息,而且平方式增长的关系配对的特征数量会给后续的推理网络在计算量和参数量上带来很大的负担。针对这个问题,提出了一种基于融合语义特征提取的引导性权重驱动的重定位关系网络模型来改善不足。方法首先通过融合场景任务的低级和高级图像特征来提取更丰富的统计图语义信息,同时提出了一种基于注意力机制的文本编码器,实现融合语义的特征提取,然后对引导性权重进行排序进一步重构图像的位置,从而构建了重定位的关系网络模型。结果在2个数据集上进行实验比较,在FigureQA(an annotated figure dataset for visual reasoning)数据集中,相较于IMG+QUES(image+questions)、RN和ARN(appearance and relation networks),本文方法的整体准确率分别提升了26.4%,8.1%,0.46%,在单一验证集上,相较于LEAF-Net(locate,encode and attend for figure network)和FigureNet,本文方法的准确率提升了2.3%,2.0%;在DVQA(understanding data visualization via question answering)数据集上,对于不使用OCR(optical character recognition)方法,相较于SANDY(san with dynamic encoding model)、ARN和RN,整体准确率分别提升了8.6%,0.12%,2.13%;对于有Oracle版本,相较于SANDY、LEAF-Net和RN,整体准确率分别提升了23.3%,7.09%,4.8%。结论本文算法围绕图表问答任务,在DVQA和FigureQA两个开源数据集上分别提升了准确率。展开更多
文摘Pulp fiber length distribution characterization has been examined in this study. Because of the fiber morphology: slender in shape, fiber size distribution characterization is a very difficult task. Traditional technique involves separation of the particles by size, such as Bauer-McNett fiber classifier, and measuring the weight fractions. The particle fractions obtained may or may not reflect the desired size classification. On the other hand, the more recent technique through optical measurement of fiber length is limited by its inability to measure the mass of the particle fractions. Therefore, not only the two techniques fail to generate identical results, either one was accepted to be of better value. Pure hardwood kraft, softwood kraft, and their mixture samples have been measured for their fiber length distributions using an optical fiber quality analyzer: FQA. The data obtained from FQA are extensively studied to investigate more reliable way of representing the fiber length data and thus examining the viable route for measuring the fiber size distributions. It has been found that the fiber length averaged length l1 is a viable indicator of the average pulp fiber length. The fiber size fraction and/or distribution can be represented by the fiber "length" fractions.
文摘目的图表问答是计算机视觉多模态学习的一项重要研究任务,传统关系网络(relation network,RN)模型简单的两两配对方法可以包含所有像素之间的关系,因此取得了不错的结果,但此方法不仅包含冗余信息,而且平方式增长的关系配对的特征数量会给后续的推理网络在计算量和参数量上带来很大的负担。针对这个问题,提出了一种基于融合语义特征提取的引导性权重驱动的重定位关系网络模型来改善不足。方法首先通过融合场景任务的低级和高级图像特征来提取更丰富的统计图语义信息,同时提出了一种基于注意力机制的文本编码器,实现融合语义的特征提取,然后对引导性权重进行排序进一步重构图像的位置,从而构建了重定位的关系网络模型。结果在2个数据集上进行实验比较,在FigureQA(an annotated figure dataset for visual reasoning)数据集中,相较于IMG+QUES(image+questions)、RN和ARN(appearance and relation networks),本文方法的整体准确率分别提升了26.4%,8.1%,0.46%,在单一验证集上,相较于LEAF-Net(locate,encode and attend for figure network)和FigureNet,本文方法的准确率提升了2.3%,2.0%;在DVQA(understanding data visualization via question answering)数据集上,对于不使用OCR(optical character recognition)方法,相较于SANDY(san with dynamic encoding model)、ARN和RN,整体准确率分别提升了8.6%,0.12%,2.13%;对于有Oracle版本,相较于SANDY、LEAF-Net和RN,整体准确率分别提升了23.3%,7.09%,4.8%。结论本文算法围绕图表问答任务,在DVQA和FigureQA两个开源数据集上分别提升了准确率。
