Open AccessMirror invariant convolutional neural networks for image classification
Author(s) -
Lu Shufang,
Li Yan,
Wang Minqian,
Gao Fei
Publication year - 2022
Publication title -
iet image processing
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.401
H-Index - 45
eISSN - 1751-9667
pISSN - 1751-9659
DOI - 10.1049/ipr2.12435
Subject(s) - convolutional neural network , artificial intelligence , computer science , invariant (physics) , pattern recognition (psychology) , transformation (genetics) , feature (linguistics) , layer (electronics) , contextual image classification , object detection , computer vision , image (mathematics) , mathematics , biochemistry , chemistry , linguistics , philosophy , organic chemistry , mathematical physics , gene
Deep convolutional neural networks (DCNNs) have been developed rapidly and they perform well on both classification and object detection tasks. However, its strong performance makes people ignore to study the invariance of the DCNNs, such as mirror invariance. In fact, the ability of DCNNs in handling mirror‐symmetrical images remains limited. In this paper, a mirror transformation convolutional layer is proposed, which transforms several feature maps to produce mirror‐symmetrical feature maps based on the traditional convolutional layer. By combining with the mirror transformation convolutional layer, the DCNNs will have mirror invariance and the performance of neural networks can be improved on classification tasks. A dataset for driver's and passenger's seatbelt detection has been collected, which is used to verify the effectiveness of the proposed convolutional layer. In the experiments, one of the state‐of‐the‐art DCNNs, GoogLeNet, is collaborated with the mirror transformation convolutional layer to form a mirror invariant networks (MINets). The experimental results show that the MINets can achieve better classification performance than the original GoogLeNet. MINets can also reduce the risk of over‐fitting caused by applying data augmentations to the dataset.