CAJ | 학술논문

针对现有方法难以解决复杂场景图像分类的问题，本文提出一种基于局部语义上下文的场景分类方法。该方法将整个图像分割为一系列超像素，从超像素提取局部特征表示图像的局部观察；在观察图像和场景类别标签之间引入表示超像素区域语义的随机变量，通过不同随机变量之间的依赖关系引入局部语义上下文信息，较好地描述了图像观察、图像内容与场景类别标签之间的语义关联度，最后定义判别图像场景类别的目标函数，采用优化方法推断图像的场景类别。在标准图像库进行的实验证明了该方法的有效性。
침대현유방법난이해결복잡장경도상분류적문제，본문제출일충기우국부어의상하문적장경분류방법。해방법장정개도상분할위일계렬초상소，종초상소제취국부특정표시도상적국부관찰；재관찰도상화장경유별표첨지간인입표시초상소구역어의적수궤변량，통과불동수궤변량지간적의뢰관계인입국부어의상하문신식，교호지묘술료도상관찰、도상내용여장경유별표첨지간적어의관련도，최후정의판별도상장경유별적목표함수，채용우화방법추단도상적장경유별。재표준도상고진행적실험증명료해방법적유효성。
An approach to scene classification based on local semantic context is proposed to solve the problem of complex scene classification. The set of super-pixel of image is segmented, and the local feature extracted from the super-pixel is used as the local image observation. The random variable is introduced to represent local semantic objects, and the spatial context of the image is described by the relationship among the variables and their neighbourhood. A new mathematical formula is used to describe the relationship among image observations, image content and scene categorization, and the semantic modeling and scene classification of the whole image is obtained. The test on the standard image dataset has demonstrated the effectiveness of the proposed approach comparable with other state-of-the-art approaches.