目的零样本汉字识别(zero-shot Chinese character recognition,ZSCCR)因其能在零或少训练样本下识别未见汉字而受到广泛关注。现有的零样本汉字识别方法大多采用基于部首序列匹配框架,即首先预测部首序列,然后根据表意描述序列(ideogra...目的零样本汉字识别(zero-shot Chinese character recognition,ZSCCR)因其能在零或少训练样本下识别未见汉字而受到广泛关注。现有的零样本汉字识别方法大多采用基于部首序列匹配框架,即首先预测部首序列,然后根据表意描述序列(ideographic description sequence,IDS)字典进行最小编辑距离(minimum edit distance,MED)匹配。然而,现有的MED算法默认不同部首的替换代价、插入代价和删除代价相同,导致在匹配时候选字符类别存在距离代价模糊和冗余的问题。为此,提出了一种字符敏感编辑距离(character-aware edit distance,CAED)以正确匹配目标字符类别。方法通过设计多种部首信息提取方法,获得了更为精细化的部首描述,从而得到更精确的部首替换代价,提高了MED的鲁棒性和有效性;此外,提出部首计数模块预测样本的部首数量,从而形成代价门控以约束和调整插入和删除代价,克服了IDS序列长度预测不准确产生的影响。结果在手写汉字、场景汉字和古籍汉字等数据集上进行实验验证,与以往的方法相比,本文提出的CAED在识别未见汉字类别的准确率上分别提高了4.64%、1.1%和5.08%,同时对已见汉字类别保持相当的性能,实验结果充分表明了本方法的有效性。结论本文所提出的字符敏感编辑距离,使得替换、插入和删除3种编辑代价根据字符进行自适应调整,有效提升了对未见汉字的识别性能。展开更多
The experimental studies of the flaring gate pier applied on the surface spillway in a high-arch dam show that a shock-wave will appear when the pattern of the flow is kept as super-critical. Meanwhile, the water dept...The experimental studies of the flaring gate pier applied on the surface spillway in a high-arch dam show that a shock-wave will appear when the pattern of the flow is kept as super-critical. Meanwhile, the water depth at the outlet increases significantly, the flow moves downward in different directions, and the plunging jet is in a narrow and long shape, with a full longitudinal diffusion. In addition, the variation of the flaring gate pier design parameters affects little the discharge capacity of the surface spillway, these parameters including the contraction ratio fl, the contraction angle c~ and the spillway chute angle O. The pressure on the bottom of the spillway increases along the way and reaches the maximum before the outlet, and then decreases rapidly. Due to the flow impacting, the pressure on both sidewalls increases abruptly at the turning line of the flaring gate pier. To see the characteristics of the flow in the flaring gate pier, a simple calculation method is suggested based on the conversation of energy and mass, and the calculation methods for the jet trajectory and the horizontal length in air are also proposed. The results are found in good agreement with experimental data.展开更多
文摘目的零样本汉字识别(zero-shot Chinese character recognition,ZSCCR)因其能在零或少训练样本下识别未见汉字而受到广泛关注。现有的零样本汉字识别方法大多采用基于部首序列匹配框架,即首先预测部首序列,然后根据表意描述序列(ideographic description sequence,IDS)字典进行最小编辑距离(minimum edit distance,MED)匹配。然而,现有的MED算法默认不同部首的替换代价、插入代价和删除代价相同,导致在匹配时候选字符类别存在距离代价模糊和冗余的问题。为此,提出了一种字符敏感编辑距离(character-aware edit distance,CAED)以正确匹配目标字符类别。方法通过设计多种部首信息提取方法,获得了更为精细化的部首描述,从而得到更精确的部首替换代价,提高了MED的鲁棒性和有效性;此外,提出部首计数模块预测样本的部首数量,从而形成代价门控以约束和调整插入和删除代价,克服了IDS序列长度预测不准确产生的影响。结果在手写汉字、场景汉字和古籍汉字等数据集上进行实验验证,与以往的方法相比,本文提出的CAED在识别未见汉字类别的准确率上分别提高了4.64%、1.1%和5.08%,同时对已见汉字类别保持相当的性能,实验结果充分表明了本方法的有效性。结论本文所提出的字符敏感编辑距离,使得替换、插入和删除3种编辑代价根据字符进行自适应调整,有效提升了对未见汉字的识别性能。
基金the National Natural Science Foundation of China (Grant Nos. 50909067, 51009102)the Program for New Century Excellent Talents in University (Grant No. 2011SCU-NCET-10-0589)
文摘The experimental studies of the flaring gate pier applied on the surface spillway in a high-arch dam show that a shock-wave will appear when the pattern of the flow is kept as super-critical. Meanwhile, the water depth at the outlet increases significantly, the flow moves downward in different directions, and the plunging jet is in a narrow and long shape, with a full longitudinal diffusion. In addition, the variation of the flaring gate pier design parameters affects little the discharge capacity of the surface spillway, these parameters including the contraction ratio fl, the contraction angle c~ and the spillway chute angle O. The pressure on the bottom of the spillway increases along the way and reaches the maximum before the outlet, and then decreases rapidly. Due to the flow impacting, the pressure on both sidewalls increases abruptly at the turning line of the flaring gate pier. To see the characteristics of the flow in the flaring gate pier, a simple calculation method is suggested based on the conversation of energy and mass, and the calculation methods for the jet trajectory and the horizontal length in air are also proposed. The results are found in good agreement with experimental data.
