Extreme learning machine with multi-structure and auto encoding receptive fields for image classification
- PDF / 2,297,536 Bytes
- 22 Pages / 439.37 x 666.142 pts Page_size
- 105 Downloads / 221 Views
Extreme learning machine with multi-structure and auto encoding receptive fields for image classification Chao Wu1 · Yaqian Li2
· Zhibiao Zhao1 · Bin Liu1
Received: 31 December 2018 / Revised: 7 December 2019 / Accepted: 9 February 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract In order to adequately extract and utilize identifiable information in the image to improve classification accuracy, extreme learning machine with multi-structure and auto encoding receptive fields (ELM-MAERF) is proposed based on local receptive fields based extreme learning machine (ELM-LRF). The ELM-MAERF is mainly composed of two convolutionpooling layers, parallel encoders and classifier. In the two convolution-pooling layers, the local receptive fields and the fully connected receptive fields are trained by utilizing the theory of ELM autoencoder. The trained receptive fields are used to extract local features, multi-channel features and fully connected features. Parallel encoders are used to adequately encode and fuse these features. The classifier trained by the approximate empirical kernel map is used to classify the fusion features, which can effectively avoid the computational difficulties caused by processing large database. To demonstrate the effectiveness of ELMMAERF, experiments are performed on four databases: Yale, MNIST, NORB and Caltech. The experimental results demonstrate the validity of trained receptive fields and structures in ELM-MAERF. Compared with the improved method based on ELM-LRF, the classification accuracy is improved by ELM-MAERF. Keywords Extreme learning machine · Local receptive fields · Fully connected receptive fields · Multi-structure · Image classification
1 Introduction Compared with the traditional gradient-based learning algorithm, the Extreme Learning Machine (ELM) proposed by Huang (2015) and Huang et al. (2004, 2006a, b) has the advantages of faster learning speed and higher generalization performance (Li et al. 2005; Wang and Han 2014). Therefore, ELM has received more and more attention from researchers, and a series of improved algorithms based on the principle of ELM have been proposed.
B
Yaqian Li [email protected]
1
School of Information Science and Engineering, Yanshan University, 066004 Qinhuangdao, China
2
Key Lab of Industrial Computer Control Engineering of Hebei Province, Yanshan University, 066004 Qinhuangdao, China
123
Multidimensional Systems and Signal Processing
Li et al. (2014) proposed a fast learning network (FLN). In the FLN, the input layer of the ELM is directly connected to the output layer. Therefore, FLN can be regarded as a combination model of linear mapping and nonlinear mapping, which effectively improves network performance. Huang et al. (2012) used the Karush–Kuhn–Tucker (KKT) theorem to train ELM and proposed KELM by introducing a kernel function. By using a kernel map for the input features, random initialization of input weights can be omitted, which can improve the stability of the network. Kasun et al. (201
Data Loading...
