Extreme learning machine with coefficient weighting and trained local receptive fields for image classification

Abstract

Local receptive fields based extreme learning machine (ELM-LRF) is widely used to solve image classification problems. However, the performance of ELM-LRF is limited by the single generation method of local receptive fields and the simple network structure. In order to solve these problems and make full use of image information to improve classification accuracy, extreme learning machine with coefficient weighting and trained local receptive fields (ELM-WLRF) is proposed based on ELM-LRF. The structure is mainly composed of convolution blocks, weighting blocks, dimensionality reduction and classification layers. In the convolution block, the principle of the ELM and the method of grouping calculation are used to train the local receptive fields of the two convolutional layers. The trained local receptive fields are used to extract identifiable feature information in the image more stably and adequately. In the weighting block, the principles of ELM and ELM autoencoder (ELM-AE) are used to train channel and spatial weighting coefficients to improve the recognizability of features. In the dimensionality reduction and classification layers, the approximate empirical kernel map (EKM) is used to train the connection weight matrix between each layer to further improve the network training speed and classification accuracy. To demonstrate the effectiveness of the proposed method, ELM-WLRF is tested on the MNIST, NORB and CIFAR-10 databases. The experimental results show that ELM-WLRF achieves superior classification accuracy, i.e. 99.27%, 98.03% and 60.14% respectively, and requires shorter training time compared with other state-of-the-art ELM-LRF-based algorithms.