LMix: regularization strategy for convolutional neural networks

• Published in: Signal, image and video processing image and video processing, 2023-06, Vol.17 (4), p.1245-1253
• Main Authors: Yan, Linyu, Zheng, Kunpeng, Xia, Jinyao, Li, Ke, Ling, Hefei
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Computer architecture; Computer Imaging; Computer Science; Computer vision; Data augmentation; Image Processing and Computer Vision; Interpolation; Masking; Masks; Multimedia Information Systems; Neural networks; Original Paper; Pattern Recognition and Graphics; Regularization; Robustness; Signal,Image and Speech Processing; Vision
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-022-02332-x

Deep convolutional neural networks perform well in the field of computer vision, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. Therefore, proper regularization strategies are needed to alleviate these problems. Currently, regularization strategies with mixed sample data augmentation perform very well, and these algorithms allow the network to generalize better, improve the baseline performance of the model. However, interpolation-based mixed sample data augmentation distorts the data distribution, while masking-based mixed sample data augmentation results in excessive information loss for overly regular shapes of masks. Although mixed sample data augmentation is proven to be an effective method to improve the baseline performance, generalization ability and robustness of deep convolutional models, there is still room for improvement in terms of maintaining the of image local consistency and image data distribution. In this paper, we propose a new mixed sample data augmentation-LMix, which uses random masking to increase the number of masks in the image to maintain the data distribution, and high-frequency filtering to sharpen the image to highlight recognition regions. We applied the method to train CIFAR-10, CIFAR-100, SVHN, and Tiny-ImageNet datasets under the PreAct-ResNet18 model to evaluate the method, and obtained the latest results of 96.32, 79.85, 97.01, and 64.16%, respectively, which are 1.70, 4.73, and 8.06% higher than the optimal baseline accuracy. The LMix algorithm improves the generalization ability of the state-of-the-art neural network architecture and enhances the robustness to adversarial samples.