基于卷积递归神经网络和核超限学习机的3D目标识别*

针对大规模RGB-D数据集中存在的深度线索质量和非线性模型分类问题,提出基于卷积递归神经网络和核超限学习机的3D目标识别方法.该方法引入深度图编码算法,修正原始深度图中存在的数值丢失和噪声问题,将点云图统一到标准角度,形成深度编码图,并结合原始深度图作为新的深度线索.利用卷积递归神经网络学习不同视觉线索的层次特征,融入双路空间金字塔池化方法,分别处理多线索特征.最后,构建基于核方法的超限学习机作为分类器,实现3D目标识别.实验表明,文中方法有效提高3D目标识别率和分类效率.

3D Object Recognition via Convolutional-Recursive Neural Network and Kernel Extreme Learning Machine

To tackle the issues of depth quality and non-linear classification in the large-scale RGB-D dataset, a 3D object recognition method is designed on the basis of convolutional-recursive neural network(CNN-RNN) and kernel extreme learning machine(KELM). Firstly, a depth coding algorithm is introduced to correct the numerical losses and noises in the original depth cue and unify the point cloud into the standard angle. And the original depth and the encoded depth are fused as the new depth cue. Secondly,multi-cue hierarchical features are learned using CNN-RNN. Meanwhile, the two-way spatial pyramid pooling method is exploited for each cue. Finally, KELM is constructed as the classifier to recognize 3D objects. The experimental results demonstrate the proposed method effectively improves the 3D object recognition accuracy and the classification efficiency.