基于RGB-D图像核描述子的物体识别方法

发布时间：2018-03-26 22:15

本文选题：RGB-D图像　切入点：物体识别　出处：《计算机应用》2017年01期

【摘要】：针对传统的颜色-深度(RGB-D)图像物体识别的方法所存在的图像特征学习不全面、特征编码鲁棒性不够等问题,提出了基于核描述子局部约束线性编码(KD-LLC)的RGB-D图像物体识别方法。首先,在图像块间匹配核函数基础上,应用核主成分分析法提取RGB-D图像的3D形状、尺寸、边缘、颜色等多个互补性核描述子;然后,分别对它们进行LLC编码及空间池化处理以形成相应的图像编码向量;最后,把这些图像编码向量融合成具有鲁棒性、区分性的图像表示。基于RGB-D数据集的仿真实验结果表明,作为一种基于人工设计特征的RGB-D图像物体识别方法,由于所提算法综合利用深度图像和RGB图像的多方面特征,而且对传统深度核描述子的采样点选取和紧凑基向量的计算这两方面进行了改进,使得物体类别识别率达到86.8%,实体识别率达到92.7%,比其他同类方法具有更高的识别准确率。
[Abstract]:The traditional methods of color depth RGB-D image object recognition have some problems, such as incomplete feature learning and insufficient robustness of feature coding, etc. An object recognition method for RGB-D image based on KD-LLC is proposed. Firstly, based on the matching kernel function between blocks, the 3D shape, size and edge of RGB-D image are extracted by kernel principal component analysis (KPCA). Color and other complementary kernel descriptors are then processed by LLC coding and spatial pooling to form corresponding image encoding vectors. Finally, these image coding vectors are fused to be robust. Simulation results based on RGB-D dataset show that as a RGB-D image object recognition method based on artificial design features, the proposed algorithm synthesizes the features of depth image and RGB image. In addition, the selection of sampling points and the calculation of compact basis vectors of traditional depth kernel descriptors are improved, which makes the recognition rate of object class reach 86.8%, and the recognition rate of entity reaches 92.7%, which is more accurate than other similar methods.
【作者单位】：武汉科技大学计算机科学与技术学院;智能信息处理与实时工业系统湖北省重点实验室(武汉科技大学);
【基金】：国家自然科学基金面上项目(41571396) 国家创新训练项目(201410488017)~~
【分类号】：TP391.41

【相似文献】