Research On RGB-D Object Recognition Algorithms Using Quaternions

Currently,some quaternion subspace analysis algorithms have been proposed.However,the performance of these algorithms is not ideal when processing nonlinear quaternion signals.Therefore,in this paper,the kernel technique is introduced to quaternion subspace analysis field.Two quaternion nonlinear subspace analysis algorithms have been proposed to process quaternion nonlinear signals.In addition,the existing quaternion representation(QR)used in quaternion subspace analysis creates redundancy and additional computation complexity when representing a color image signal of three components by a quaternion matrix having four components.The depth information is introduced into the QR to represent the RGB-D images.This new QR fully utilizes the four-dimensional quaternion domain.Finally,this paper combines the new QR with quaternion nonlinear subspace analysis algorithm for RGB-D object recognition.The main research works are as follows:(1)Propose a new QR for RGB-D images.The important depth information is considered to the existing QR for representing the RGB-D images.(2)Propose two quaternion nonlinear subspace analysis(QNSA)algorithms.Based on the existing quaternion principal component analysis(QPCA)and quaternion linear discriminant analysis(QLDA),both kernel quaternion principal component analysis(KQPCA)and quaternion generalized discriminant analysis(QGDA)are proposed by introducing the kernel technique into quaternion subspace analysis.Moreover,using the block-based idea and the parallel computing idea for KQPCA,the block-wise 2D KQPCA(B2DKQPCA)is proposed to implement 2D KQPCA really.(3)Study the application o.f QNSA algorithms in RGB-D object recognition field.The QNSA algorithms are applied into RGB-D object recognition by combining with the new QR for RGB-D images in the work(1).Experimental results on three public datasets(?I-D dataset,CIN2D/3D dataset and FERET dataset)demonstrate that the proposed RGB-D object recognition algorithms using QNSA algorithms achieve better performance than some algorithms based on quaternion linear algorithms and some CNN-based algorithms.In addition,the B2DKQPCA-based algorithm achieves the best performance among the algorithms using three proposed QNSA-based algorithms.