数据驱动的人体图像语义分割研究

项目名称： 数据驱动的人体图像语义分割研究

项目编号： No.61402430

项目类型： 青年科学基金项目

立项/批准年度： 2015

项目学科： 自动化技术、计算机技术

项目作者： 王丹

作者单位： 中国科学院大学

项目金额： 10万元

中文摘要： 人体图像的语义分割，即按照头发、肤色、上身衣服、下身衣服、鞋袜、背景等类别将人体图像进行像素或超像素级的语义标注。该问题涉及统计学习、图像处理等多个学科领域，具有重要的理论价值，同时它可以直接应用于行人身份识别、人体服饰检索、人体动作识别等实际问题。为此，本项目拟开展的主要研究内容包括： 1)人体图像的高效标注，拟采用主动学习的策略，快速获取大量具有语义分割的训练数据；2)人体定位，即确定整个人体区域的大体位置以排除大部分背景区域，拟采用显著性建模方法进行预滤波以提高定位的速度和精度；3)人体部件的形状建模，拟采用数据驱动的方法，自适应地学习针对特定实例的可形变形状模型，并将该高层形状模型与低层表观模型结合，从而精确分割人体图像中的各个部件区域。本项目有望用于电子商务的服饰分割、检索，以及智能视频监控的人体身份识别问题中。

中文关键词： 图像语义分割；人体图像分割；深度学习；显著性检测；

英文摘要： Semantic segmentation for human images is to label each pixel as one of the categories including hair, skin, upper body clothing, lower body clothing, footwear, background or others. It is related to multiple disciplines of statistical learning, image processing and other fields and thus is worth studying. Meanwhile, it can directly facilitate other practical tasks, such as human identification, human clothing retrieval and human action recognition. To this end, project will study the following contents: 1) Efficient labeling of human images. We plan to utilize active learning based strategy, to quickly get a lot of training data. 2) Locating human body, which determines an approximate position of the entire area of the body, while excluding part of the background area. We intend to adopt saliency modeling as a filtering step, which can improve the speed and accuracy of human location. 3) Building the shape model for each body part. Data-driven method will be explored to adaptively learn instance-specific deformable shape model. Moreover, the high-level shape model will be integrated with low-level appearance models to achieve accurate segmentation of human parts. This project is expected to dress for the e-commerce clothing segmentation, retrieval and human recognition in intelligent video surveillance.

英文关键词： image semantic segmentation；person segmentation；deep learning；saliency detection；