Page 319 - 《软件学报》2025年第7期

3240 软件学报 2025 年第 36 卷第 7 期

和主题, 可能涉及多个标签, 而传统的单标签文本分类方法只能为每个文本分配一个最相关的标签. 因此, 在处理
涉及多主题的文本数据时, 多标签文本分类是一种更有效的方法.
多标签文本分类任务是为每个文本分配多个标签, 与传统的单标签文本分类任务不同, 多标签文本分类面临
着更大的挑战和复杂性, 如标签依赖 [6] 、标签不平衡 [7] 、标签噪声 [8] 等. 早期的研究方法将多标签分类问题转化为
多个独立的单标签分类问题, 针对每个标签独立训练一个分类器来预测该标签, 然后将所有分类器的输出组合成
最终的多标签预测结果 [9] . 这类方法简单易实现, 但忽略了标签之间的关系. 随着深度学习的不断发展, 基于深度
神经网络的方法被陆续提出, 这类方法主要通过提取增强文本表示和建模标签关系来改善分类效果. 在提取增强
文本表示上, 深度学习模型通过学习文本的语义表示和上下文信息, 能够更好地捕捉文本中的关键信息, 从而提高
分类的性能 [10,11] . 在建模标签关系上, 利用标签的先验知识 (如共现等) 构造标签图, 采用不同的方法在图上学习标
签表示, 通过捕获标签之间的依赖关系提升分类性能 [12,13] . 在这些方法当中, 标签图通常基于数据集统计信息构
建, 也即是静态图. 然而, 从训练集获取到的标签共现数据依赖全局数据分布, 不能代替文档内容特有的标签关系.
在实际情况中, 不同文本依赖不同的标签关系, 需要一种动态的方式描述标签关联.
图 1 展示了动态建图与静态建图的差异. 为文本中的词汇标注标签, 并根据标签共现关系动态调整标签关系图.

cs.MA
cs.AI cs.RO cs.CV cs.AI
cs.CV
intelligent machines require basic information such as moving object detection from
cs.AI
videos in order to deduce higher level semantic information . in this paper , we
cs.CV cs.AI
propose a methodology that uses a texture measure to detect moving objects in vi
deo the methodology is computationally inexpensive , requires minimal parameter fi 静态图
-ne tuning and also is resilient to noise , illumination changes , dynamic backgroun cs.RO cs.SY
d and low frame rate experimental results show that performance of the proposed
cs.MA
approach is higher than those of state of the art approaches we present framework
cs.AI cs.SY cs.MA cs.CV cs.AI cs.CV
for vehicular traffic density estimation using foreground object detection technique
cs.AI
动态图
cs.CV cs.AI
and present comparison between the foreground object detection based framework
cs.AI cs.SY cs.MA cs.AI cs.SY cs.MA
and the classical density state modelling based framework for vehicular traffic densi
ty estimation cs.RO cs.SY
(a) 动态图样例 1: 捕捉静态图中不存在的标签关联
cs.MA
cs.CV cs.AI
cs.CV
texture classification became one of the problems which has been paid much attention on
cs.CV cs.AI cs.AI
by image processing scientists since late 80s consequently , since now many different
methods have been proposed to solve this problem in most of these methods the
researchers attempted to describe and discriminate textures based on linear and non linear 静态图
patterns the linear and non linear patterns on any window are based on formation of
cs.CV cs.AI cs.RO cs.SY
grain components in a particular order grain component is a primitive unit of morphology
that most meaningful information often appears in the form of occurrence of that the cs.MA
approach which is proposed in this paper could analyze the texture based on its
cs.CV
cs.CV cs.AI cs.CV cs.AI
grain components and then by making grain components histogram and extracting 动态图 cs.AI
statistical features from that would classify the textures finally , to increase the accuracy of
classification , proposed approach is expanded to color images to utilize the ability of
approach in analyzing each rgb channels , individually although , this approach is a general
one and it could be used in different applications , the method has been tested on the
stone texture and the results can prove the quality of approach cs.RO cs.SY
(b) 动态图样例 2: 削弱标签关系图中文本无关标签的关联
图 1 动态图的生成

图 1(a) 中描述的文档中包含 detect moving objects、moving object detection、vehicular traffic density estimation、
foreground object detection 等关键词, 这些关键词不仅与人工智能 (cs.AI) 和计算机视觉 (cs.CV) 相关, 也与系统控