基于多模态深度学习的RGB-D物体识别

发布时间：2018-05-10 10:14

本文选题：多模态 + 稀疏自编码　；参考：《计算机工程与设计》2017年06期

【摘要】：针对现有RGB-D物体识别方法存在图像特征学习不全面、类间相似的物体识别精度不高等问题,联合稀疏自编码(sparse auto-encoder,SAE)及递归神经网络(recursive neural networks,RNNs)提出多模态稀疏自编码递归神经网络(multi-modal sparse auto-encoder and recursive neural networks,MSAE-RNNs)的深度学习模型。SAE结合卷积及池化技术分别从RGB-D图像的RGB图、灰度图、深度图、3D曲面法线中提取低层次的平移不变性特征,作为多个固定树RNNs的输入,得到更加抽象的高层特征,融合后的多模态特征,采用SVM分类器进行分类。在RGB-D数据集上的实验结果表明,该算法的物体识别率达到90.7%,较其它算法提高了3%-9%,能很好完成RGB-D物体的识别。
[Abstract]:In view of the existing RGB-D object recognition methods, the image feature learning is not comprehensive, and the recognition accuracy of objects similar to each other is not high. Combined sparse self-coding Sparse auto-encodern (SAE) and recursive neural network recursive neural networks (RNNs), a multi-modal sparse self-coding recursive neural network multi-modal sparse auto-encoder and recursive neural networks-MSAE-RNNs is proposed. SAE combines convolution and pool techniques from RGB images, grayscale images, gray images of RGB-D images, etc. The translation invariance feature of low level is extracted from 3D surface normals of depth map, which is used as input of multiple fixed tree RNNs to obtain more abstract high-level features and multimodal features after fusion. SVM classifier is used for classification. The experimental results on the RGB-D dataset show that the recognition rate of the proposed algorithm is 90.7, which is 3- 9 higher than that of other algorithms, and the recognition of RGB-D objects can be accomplished well.
【作者单位】：武汉科技大学计算机科学与技术学院;武汉科技大学湖北省智能信息处理与实时工业系统重点实验室;
【基金】：国家自然科学基金面上基金项目(41571396) 国家创新训练基金项目(201410488017) 省级大学生创新创业训练计划基金项目(201410488038) 校级大学生科技创新基金项目(14ZRA079、14ZRC093)
【分类号】：TP183;TP391.41

【相似文献】