Research on Image Classification Improvement Based on Convolutional Neural Networks with Mixed Training
Yongyue Zhang, Junhao Zhang, Wenhao Zhou · 2022 IEEE 4th International Conference on Civil Aviation Safety and Information Technology (ICCASIT) · 2022
Convolutional neural networks (CNNs) have been widely used and have shown excellent predictive power in image classification. In this paper, an ensemble method, namely the Mixture of trained CNNs (M-tCNNs) method, was proposed to improve the classification performance evaluated on the CIFAR-100 dataset. The M-tCNNs consists of two components, including three expert networks (CNNs) and a trainable gating network. The classification performance of M-tCNNs method was compared to that of single CNN and simple average (SA) method. The results showed that the M-tCNNs method achieved a better accuracy (ACC) of 84.18% compared to single CNN (highest ACC = 78.74%) or SA method (ACC = 80.78%). Experimental results indicate that M-tCNNs method can improve the classification performance for CIFAR-100 dataset compared to single CNN and SA methods.