Title: The image classification algorithm research using class information loss and joint structural similarity
Authors: Shian Wang
Addresses: Department of Information Engineering, Guangzhou Institute of Technology, Guangzhou, Guangdong, 510075, China
Abstract: Aiming at the supervised training of convolutional neural networks, the weighted joint structural similarity and class information supervised training method has been proposed. Firstly, for a small image, the convolutional neural networks that can extract high-level information of images is designed. Secondly, a weighted joint structural similarity and class information loss function training convolutional neural network are established. Finally, handwritten numbers and Cifar10 images are obtained by Mnist dataset. The image classification experiments can validate the effectiveness of the proposed network. The experimental results can show that the image classification error rate of improved network on Mnist handwritten digits and Cifar10 dataset is 0.23% and 10% respectively. Under the premise that there is no dataset increase on the Mnist dataset, the performance of proposed network exceeds the performance. The performance of all single networks on the dataset, on the Cifar10 dataset, the proposed network can achieve higher image classification accuracy with less computational effort. At the same time, the supervision of joint structural similarity and class information loss can speed up the training process of proposed network.
Keywords: convolutional neural network; image classification; structure similarity; deep learning; metric learning.
DOI: 10.1504/IJICT.2020.106313
International Journal of Information and Communication Technology, 2020 Vol.16 No.3, pp.203 - 213
Received: 23 Nov 2018
Accepted: 01 Apr 2019
Published online: 02 Apr 2020 *