RGB-D Image Recognition Via Deep Neural Networks

With the development of depth acquisition technology,combining RGB and depth features has become a hot research in the field of image recognition which is widely used in robotic navigation,aerospace and virtual reality.To solve the issues of depth quality,non-linear classification and image occlusion in the large-scale RGB-D dataset,we present RGB-D image recognition based on deep neural network from the third-person and first-person views.The main work of this thesis is illustrated as follows:(1)We introduce the research background of RGB-D image recognition,status and difficulties.Then we introduce depth acquisition method,and subsequently summarize RGB-D image representation methods from the aspects of hand-carfted features and deep learning.Finally we induce three kinds of classifiers.(2)We design and achiecve a third-person RGB-D image recognition method via Convolutional-Recursive Neural Network(CNN-RNN)and Kernel Extreme Learning Machine(KELM).Firstly,we introduce a depth coding algorithm to correct original depth and fuse them as the new depth cue.Then,we use CNN-RNN to learn multi-cue hierarchical features.Meanwhile,we exploit the two-way pyramid pooling method to unify the feature size.Finally,we construct KELM as the classifier.Experimental results show that the method significantly increases the classification accuracy and efficiency.(3)We design and achiecve a first-person RGB-D image recognition method based on multi-modal feature fusion to recognize grasp types.Firstly,we utilize depth images to extract hand features and effectively eliminate image occlusion.In order to fully exploit the interrelationship between RGB clue and depth clue,we exploit a CNN model via multi-modal feature fusion algorithm for multi-feature learning.Finally,we construct KELM as the classifier to accomplish recognition task.Experimental results show that the method can improve the recognition performance to some extent.