Research Of RGB-D Images Recognition Algorithms Based On Deep Learning Network

3D images classification is one of hot researches in the field of computer vision,It is widely used in indoor object detection,pedestrian detection,attitude detection and robot vision.With the rapid development of artificial intelligence,RGB-D images classification based on deep learning has been widely concerned.We design and achieve a method of RGB-D images classification based on deep learning network with sparse connections for extracting features effectively.As for the issues of large weight parameter that occupy larger memory space in deep learning network,We achieve a method of weight sharing each layer based on deep convolutional neural network.The main contribution is as follows:(1)The thesis expounds the research background and meaning of RGB-D images classification based on deep learning,and summarizes its research status.We introduces the correlation theory about images classification and deep learning network emphatically.We analyze methods of collecting depth images,feature extraction about RGB images and depth images,classical image classification models and classical deep learning networks models in detail.(2)In order to extract targeted features of RGB-D images automatically,We design and achieve a method of RGB-D images classification based on deep learning network with sparse connections.In the filters layer of convolutional neural network,we get the image feature points through SURF algorithm,and obtain the receptive field by the center of feature points.Connecting the receptive field and convolution kernels to form a sparse connections network,to form the convolutional neural network of filters layer.The experimental results show that the proposed method can effectively improve the recognition accuracy and is robust.(3)As for the issues of expensive amount of computation,large weight parameter and memory consumption in deep learning network,In this thesis,We design and achieve a method of RGB-D images classification based on deep convolutional neural network with weight sharing each layer.Clustering weight of each layer network by SOM algorithm,then achieving single weight sharing by the center of cluster.The experimental results show that the deep learning network can be effectively compressed by this method with the RGB-D images recognition rate remain unchanged approximately.