【摘要】 本文提出了一种新型的结合超限学习机(ELM)和融合卷积网络(CCN)的模型,并用于3D物体的特征提取和分类.模型以3D物体的多视角投影图作为输入,经过多层融合卷积网络提取特征,利用半随机的ELM网络进行分类.卷积网络由提出的融合卷积单元组成,它是一种改进的残差单元,多个并行残差通道上的卷积核个数依次增加,相同大小的卷积核参数共享.半数卷积核参数以高斯分布随机产生,其余通过训练寻优得到.它能拟合更复杂的残差项函数,增加低层网络的特征表达能力.同时网络结构规范简洁,便于训练和优化.本文的方法在普林斯顿3D模型标准数据集上的识别率达到了92. 86%.实验表明,提出的方法的识别率比现有的ELM方法和深度学习等最新方法的识别率更高,并且其调节参数少,收敛速度快.更多还原

【Abstract】 This paper proposes a new model that combines Extreme Learning Machine( ELM) and Coalesce Convolutional Network( CCN) for feature extraction and classification of 3D objects. This model takes multi-view projections of 3D objects as input,extracts the features through multi-layer coalesce convolution network,and classifies them by semi-random ELMnetwork. The convolutional network consists of many coalesce convolution blocks that proposed in this paper,and each of them is an improved residual unit. The number of convolution kernels on multiple parallel residual channels increases sequentially,and convolution kernel parameters of the same size are shared. The half convolution kernels’ parameters are randomly generated by Gaussian distribution,and the rest are obtained by training optimization. It can fit more complex residual functions and increase the characterization ability of low-level networks. At the same time,the network structure is more standardized and simple,and it is easy to train and optimize. The recognition accuracy of proposed method in the Princeton 3D model benchmark dataset reaches 92. 86%. Experiments show that the proposed method has higher recognition accuracy than the latest ELMmethods and deep learning methods. Its adjustment parameters are less and the convergence speed is fast.更多还原