【摘 要】
：
针对现有火焰图像分割精度低的问题,提出了一种基于改进ResNet-UNet的火焰图像分割网络,实现火焰的精确分割。以UNet神经网络为基础,采用ResNet-34网络作为神经网络特征提取前端,加深了网络层数,保留ResNet-34的独立卷积结构和残差结构,并保证ResNet-34网络和UNet网络的有效融合,从而形成ResNet-UNet神经网络。针对火灾图像中目标区域与背景区域在图像中所占比例差
【机 构】
：
北京信息科技大学机电工程学院,北京100192
【出 处】
：
北京信息科技大学学报(自然科学版)
论文部分内容阅读
针对现有火焰图像分割精度低的问题,提出了一种基于改进ResNet-UNet的火焰图像分割网络,实现火焰的精确分割。以UNet神经网络为基础,采用ResNet-34网络作为神经网络特征提取前端,加深了网络层数,保留ResNet-34的独立卷积结构和残差结构,并保证ResNet-34网络和UNet网络的有效融合,从而形成ResNet-UNet神经网络。针对火灾图像中目标区域与背景区域在图像中所占比例差别大而导致的分割类别不平衡问题,将交叉熵损失与Dice损失线性组合并对Dice损失加权。经实验证明,提出
其他文献
I am a gardener working in the garden. The garden is teaching me about working with the earth. I recognize that there will be both successes and failures and there are many variables that affect them both.rnThe quality of the seeds planted has a bearing o
I grew up in a small town where the elemen-tary school was a ten-minute walk from my house. When we children could go home for lunch, we would find our mothers waiting. At the time, I did-n\'t consider this to be so important, although today it certainl
1 There are 56 different ethnic groups in China. Therefore, Chinese attire also includes various clothing worn by different Chinese ethnic groups. Here are a few examples:rn·Dai ethnic grouprn2 Although the Dai are amongst the 56 ethnic groups of China, t
1 Each year, the world wastes about one-sixth of the food available to consumers. It is estimated that the world has lost about 931 million metric tons of food so far. That\'s about an average of 121 kilograms for each man, woman and child on Earth.rn2
An activity about the short English dramas kicked off in Yuqi Senior High School. The activity was meant to show students\' talents for acting out fantastic English dramas, ranging from classics to modern ones.rnTo begin with, students were supposed to
After I had left the laboratory, I escaped into the country outside the town. I soon felt hungry and thirsty, and my first food was fruit which I found on some trees near a river. I drank from the river and then lay down and went to sleep.rnAt first, my e
针对大规模数据场景下K-means聚类算法存在总体速度较慢的问题,提出了一种基于大数据流计算框架Apache Flink,并结合流处理思想的SK-means策略.从算法优化方面,采用流处理的思想在数据读入的过程中确定聚类的初始聚类中心;从实现方案上,选用Apache Flink框架设计了一种基于SK-means策略的优化方案,分析该策略对聚类效果和耗时的影响.实验结果表明:该策略在数据量较大的场景中有较好的优化效果,不需要预先假设出聚类数量和初始聚类中心的具体值,有效地降低了聚类迭代次数,降低了整体计算耗
为提升政府公文写作的效率,设计开发了一种公文智能辅助写作系统,在机关单位人员撰写公文时可检索、推荐相似的政府公文以作参考.使用Lucene对政府公文进行倒排索引,实现了公文的关键词检索.提出了一种多因素加权融合相似度计算算法,将公文标题的编辑距离和公文文本的语义距离加权融合.系统在THUCNews数据集上的结果优于TF-IDF和Word2Vec.目前已经应用于私有云环境下服务化智能办公平台,具有较好的实用性.
One day in summer, a lady on the horseback rode up to the house and knocked on the door. She had dark hair and was very beautiful. The family were all very pleased to see her. I soon learned that her name was Sophie.rnShe could not speak the family\'s l
在实体关系抽取和分类的任务背景下,针对当前基于深度学习模型对实体间语义关系利用不足的问题,提出了基于预训练ALBERT模型的实体关系分类模型.首先使用实体定位符号对实体位置进行定位,然后使用依存关系方法提取实体间关系描述词,并利用预训练ALBERT模型对输入进行向量化表征,再利用注意力机制将向量进行融合,最终使用Softmax分类器得到关系分类结果.与其他几种深度模型相比,该模型在KBP37数据集和Semeval-2010-task8数据集上的结果最优,验证了模型的有效性.