融合RGB-D信息的三维物体识别算法

Three-dimensional object recognition algorithm fusing RGB-D information

图像处理是物体识别的关键环节,不同的模态特征之间具有互补性,同时使用能够提高目标的识别准确率,但现有研究仅仅是将多模态特征直接融合或者人工构造特征描述子进行识别工作,没有区别对待不同模态的不同特征且忽略了特征的内部联系。为了更客观地反映物体三维特性,结合稀疏自编码网络和改进的卷积神经网络,提出一种新的深度学习模型SAE-RCNN与一种分段训练网络的方法,可以提取有辨别力的特征而且避免了网络退化的问题,并将特征在全连接层高效融合,通过分类器Softmax得到实验结果。实验数据采用Washington RGB-D标准数据集。结果表明,SAE-RCNN算法模型的物体识别率达到89.7%,较其他算法取得了更好的识别效果。

Object recognition is a key link of image processing.The complementarily of different modal features can improve the recognition accuracy of objects.However,the existing research only involves the fusion of multimodal features or the construction of feature descriptors,but does not treat the different features discriminatively and ignores the internal relation of the features.In order to reflect the three-dimensional features of objects more objectively,a new deep learning model SAERCNN(sparse autoencoder-region with CNN features)and a multi-stage network training method are proposed in combination with the sparse self-coding network and the improved convolutional neural network.The algorithm can extract features with discriminative power while avoiding network degradation and fuse them efficiently in the fully connected layer.The experimental result was obtained with classifier Softmax.The Washington RGB-D standard database is adopted for the experiment data.The results show that the object recognition rate of SAE-RCNN algorithm model reaches 89.7%,which is better than other algorithms.