Research On RGB-D Object Recognition Technology Based On Deep Learning

Object recognition gradually plays a considerable role in daily life.Bank identification,mobile recognition,entrance guard and industry 3D reconstruction use related technologies to work efficiently.The traditional method of object recognition based on labeling is demanding,so the pursuit of the rapid detection has become a trend in today.In the past few years,many researchers combine machine learning with recognition field,which produces huge impact on object recognition,improving the availability of technologies and generating great value.This paper improves the network structure and comes up with a new network model to settle a matter of low recognition efficiency,poor real-time performance and declining network performance existing in previous studies,so as to make accuracy be better.First,this paper makes the process of general object recognition clear and explains the access to information,feature extraction,feature classification and key points in CNN.To address the low resolution of depth map,a new reconstruction algorithm is raised based on SRCNN network model by using the guidance of color information with depth information.Compared with the traditional image reconstruction algorithm,the image quality obtained by the algorithm in this paper is significantly improved.Then,a method of feature fusion between RGB model and depth model is proposed.Due to the correlation between different models of the same object,the ratio comparison of RGB model and depth model can be obtained by using the trained single stream CNN under the powerful learning ability of neural network,the relative importance of different models can be used by the medium-term fusion method,improving the accuracy of object recognition.Finally,the SAE-RCNN network model is proposed on the basis of clarifying the relative importance of different models by combining the sparse self-coding network and the improved CNN.The test is carried on the Washington RGB-D dataset.Using the SAE-RCNN algorithm model,RGB model and depth model features are combined by middle fusion and sent to residual network for learning to get the precision.The experimental results show that the proposed model by this paper achieves better accuracy than others.