基于ResNet的RGB-D物体识别

RGB-D Object Recognition Based on ResNet

摘要: 传统的物体识别研究一般是基于RGB图像和灰度图像，RGB图像和灰度图像自身具有一定的局限性，由于缺少物体表面的形状信息，对颜色相近但形状不同的物体进行识别时容易造成错误。使用RGB-D相机可以在获取高分辨的RGB图像的同时，获取每个像素的深度值，深度数据中蕴含着有关物体形状的信息，可以为物体识别提供新的特征。本文主要研究了基于深度学习的RGB-D物体识别算法，选择了ResNet残差网络的V1b型，通过手动调参设置适合于提取RGB-D图像特征的卷积神经网络模型。通过两个ResNetV1b网络模型分别提取有效的RGB特征和深度特征，为了提取到更多样性的特征，在残差网络子模块中加入了空洞卷积；采用了全连接的方法对RGB特征和深度特征进行融合。实验证明了本文所采用的神经网络结构对于RGB-D物体识别是有效的，对比实验表明空洞卷积的引入有效提高了RGB-D图像的识别率。

Abstract: Traditional object recognition research is generally based on RGB image and gray image. RGB image and gray image have their own limitations. Due to the lack of shape information of the object surface, it is easy to make mistakes when recognizing objects with similar color but different shapes. Using RGB-D cameras, we can obtain the depth value of each pixel while obtaining high-resolution RGB images. The depth data contains the information about the shape of the object, which can provide new features for object recognition. This paper mainly studies the RGB-D object recognition algo-rithm based on deep learning, selects the V1B type of ResNet residual network, and sets the convolution neural network model suitable for extracting RGB-D image features through manual parameter adjustment. Two ResNet V1b network models are used to extract effective RGB features and depth features respectively. In order to extract more diverse features, dilated convolution is added to the residual network. A full connection method is used to fuse RGB features and depth features. Experiments show that the neural network structure used in this paper is effective for RGB-D object recognition, and the contrast experiments show that the introduction of dilated convolution effectively improves the recognition rate of RGB-D images.