基于有监督学习的自然图像中骨架提取和物体识别研究

项目名称： 基于有监督学习的自然图像中骨架提取和物体识别研究

项目编号： No.61303095

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 沈为

作者单位： 上海大学

项目金额： 25万元

中文摘要： 在自然图像中实现物体的表示与识别是计算机视觉领域的难点问题。本项目研究基于有监督学习的自然图像中骨架提取和基于所提取骨架的图像中物体识别。研究内容包括：研究基于学习的自然图像中骨架检测算法，主要解决对称区域特征描述符的构造问题；研究如何根据检测得到的骨架重建准确的区域边界并同时对骨架位置进行求精，解决该协同求精中涉及的多变量方程迭代优化问题；研究如何根据重建区域特征对骨架进行去噪，解决噪声骨架的区分性特征的设计问题；在根据骨架段的几何一致性进行骨架段连接后，研究基于提取骨架结合表象模型的图像中物体定位识别算法，解决定位中涉及的骨架重建区域融合的组合优化问题。最后，以本项目方法模型和理论算法为基础，研究其他相关应用问题，如文本检测与识别和行人检测等。

中文关键词： 骨架；物体检测；边缘检测；有监督学习；文本检测

英文摘要： Object representation and recognition in natural images is a challenging problem in Computer Vision. The project researches supervised learning based skeleton extraction in natural images and object recognition via the extracted skeletons. The research contents include: the research on supervised learning based skeleton detection in natural images, in which we mainly address the problem that how to form an informative symmetric region characteristic descriptor; the research on how to reconstruct the accurate region boundary from the detected skeleton and refine the positions of the skeleton synergistically, in which we address the problem of multivariable function optimization involved in such a co-refinement; the research on how to perform skeleton denoising according to the reconstructed region, in which we address the problem that how to extract discriminative feature to classify skeletons and noises; the research on object localization and recognition based on the extracted skeletons and the appearance model after skeleton linking with geometry consistency, in which we address the problem of combinatorial optimization to merge the regions reconstructed from the extracted skeletons for object localization. Some other problems related to this project (eg. Text detection and recognition, pedestrian detection, a

英文关键词： Skeleton；Object Detection；Edge Detection；Supervised Learning；Text Detection