Real-World Object Recognition for Robot Based on Lightweight Deep Multimodal Distance Learning

Abstract

Real-world object recognition is an important and difficult robot vision problem. In this paper, a real-world multi-angle and multi-attitude deformable object recognition method for robot system, named RCOR, is proposed based on lightweight deep multimodal distance learning (DMDL). (1) Deep multimodal convolutional neural network (DMCNN) is proposed to improve the transformation abilities of CNNs and enhance feature maps’ resolutions. (2) Deep distance metric learning (DDML) is presented to relieve the problem of lacking adequate labeled data and efficiently reduce redundancy. (3) To apply RCOR into embedded vision applications in real-world environment, a light weight DCNN, Mobile-XB, is proposed. Extensive experiments demonstrate that the proposed approach significantly outperforms state-of-the-arts. And it performs well on computationally limited platforms.