RGB-D Object Recognition Method Based On Tensor Decomposition And Convolutional Neural Network

Object recognition is an important research direction in the field of computer vision.Its main purpose is to enable computers to “see” the objects in the real world and allow computers to have the ability to perceive objects.In the era of rapid proliferation of image data,the ability of computers to automatically recognize objects in images will greatly improve people's work efficiency in processing images,excavating images,and managing images,and thus has a huge application prospect.The appearance of the depth sensor enables a new generation of cameras to acquire RGB-D images of objects,where the depth image compensates for missing spatial structure information of the objects in the RGB image.How to combine the RGB-D image with the object recognition technology effectively and improve the accuracy of object recognition has become a new research hotspot in the field of computer vision.This thesis focuses on the RGB-D object recognition technology and focuses on the RGB-D image fusion and prediction model construction problems.The RGB-D object recognition method based on tensor decomposition and convolutional neural network is studied.An RGB-D image fusion method based on tensor decomposition and a convolution kernel number determination method based on edge detection are proposed.The main work and contributions of this article are as follows:(1)The overall framework of RGB-D object recognition based on tensor decomposition and convolutional neural network is designed.The framework includes RGB-D image fusion and prediction models to build two modules.The RGB-D image fusion module mainly includes image preprocessing and tensor decomposition;the convolutional neural network model is used as a prediction model in the prediction model building module(2)Aiming at the problem of how to effectively use the spatial structure information in depth image to improve the recognition accuracy in RGB-D object recognition,an RGB-D image fusion method based on tensor decomposition is proposed.This method mainly draws on the advantages of tensor decomposition,and creatively uses tensor decomposition to solve RGB-D image fusion problems.This method first analyzes the properties of RGB-D image data,and then constructs the corresponding tensor.Then,Tucker method is used to decompose the tensor to obtain the factor matrix.Finally,the original tensor is projected by the factor matrix to obtain the fused RGB.-D image.The simulation experiment results show that this method can improve the accuracy of object recognition and can increase 19,7 and 11 percentage points respectively in the three sub-data sets.(3)Aiming at the problem that the number of convolution kernels in the construction of convolutional neural network prediction model is determined by experience,a method for determining the number of convolution kernels based on edge detection is proposed.The method firstly performs edge detection on the training image to obtain the edge image,then extracts the edge block of the edge image and the statistically extracted edge block is stored in the edge feature matrix.Finally,the edge feature matrix is analyzed to obtain the number of convolution kernels.The experimental results on the Mnist data set and the Chars74 K data set show that the method can adaptively increase or decrease the number of convolution kernels according to the image set and has the ability to adapt.