摘要： RGB图像和深度图像的同时使用能有效提高物体识别的准确率。然而,已有研究仅将RGB图像和深度图像的特征进行简单的线性连接,没有根据RGB特征和深度特征的差异性进行特征提取和融合,充分发挥RGB-D图像的优势。为此,提出一种多模态稀疏自编码算法,在进行差异性特征提取的同时完成RGB特征和深度特征的有效融合。结合多模态稀疏自编码算法和空间金字塔最大池化算法,给出一个全新的深度学习模型。该模型能够提取有辨别力的特征并完成基于RGB-D图像的物体识别工作。在2个标准的RGB-D数据库上的实验结果表明,与基于RGB-D的物体识别算法相比,该算法能够有效融合RGB特征和深度特征,取得更高的识别准确率。

关键词: RGB特征与深度特征融合, 稀疏自编码, 多模态稀疏自编码, 空间金字塔最大池化, 深度学习, 物体识别

Abstract: Combining RGB image and depth image can effectively improve the RGB-D image recognition accuracy.However,prior researchers only do simple linear connect with the RGB image and depth features and do not extract and fuse the RGB and depth features according to their difference,and do not take full advantage of RGB-D image.This paper proposes a multi-model sparse auto encoder algorithm.Multi-model sparse auto encoder algorithm can extract and fuse the RGB and depth features at the same time.By combining multi-model sparse auto encoder algorithms with spatial pyramid max pooling algorithms,it proposes a new deep learning model.New depth learning model can extract recognizable features and complete the RGB-D based object recognition.It uses two standard RGB-D databases to verify the new proposed algorithm and deep learning model.Experimental results show that compared with previous RGB-D image based object recognition algorithm,the newly proposed algorithm effectively fuses the RGB and depth features and achieves higher recognition accuracy.

Key words: RGB feature and depth feature fusing, Sparse Auto Encoding(SAE), Multi-model Sparse Auto Encoding(MMSAE), spatial pyramid max pooling, deep learning, object recognition

中图分类号: 

• TP391.06