基于RGB-D图像核描述子的物体识别方法

Object recognition method based on RGB-D image kernel descriptor

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。

The traditional RGB-Depth （RGB-D） image object recognition methods have some drawbacks, such as insufficient feature learning and poor robustness of feature coding. In order to solve these problems, an object recognition method of RGB-D image based on Kernel Descriptor and Locality-constrained Linear Coding （KD-LLC） was proposed. Firstly, based on the kernel function of image block matching, several complementary kernel descriptors from RGB-D images, such as 3D shape, size, edges and color, were extracted using Kernel Principal Component Analysis （KPCA）. Then, the extracted feature from different cues, were processed by using LLC and Spatial Pyramid Pooling （SPP） to form the corresponding image coding vectors. Finally, the vectors were combined to obtain robust and distinguishable image representation. As a hand- crafted feature method, the proposed algorithm was compared to other hand-crafted feature methods on a RGB-D image dataset. In the proposed algorithm, multiple cues from depth image and RGB image were used, and the sampling points selection and basis vectors calculation schema for depth kernel descriptor generation were proposed. Due to above-mentioned improvements, the category and instance recognition accuracy of the proposed algorithm for objects can respectively reach 86.8% and 92.7%, which are higher than those of the previously hand-crafted feature methods for object recognition from RGB-D images.