MeshCut data augmentation for deep learning in computer vision.

Wei Jiang,Kai Zhang,Miao Yu,Nan Wang,Zhishun Wang

doi:10.1371/journal.pone.0243613

Wei Jiang, Kai Zhang + Show 3 more

Open Access

https://doi.org/10.1371/journal.pone.0243613

Copy DOI

Journal: PloS one	Publication Date: Dec 23, 2020
Citations: 12	License type: CC BY 4.0

Affiliation: Sichuan University, Southwest Jiaotong University

Abstract

To solve overfitting in machine learning, we propose a novel data augmentation method called MeshCut, which uses a mesh-like mask to segment the whole image to achieve more partial diversified information. In our experiments, this strategy outperformed the existing augmentation strategies and achieved state-of-the-art results in a variety of computer vision tasks. MeshCut is also an easy-to-implement strategy that can efficiently improve the performance of the existing convolutional neural network models by a good margin without careful hand-tuning. The performance of such a strategy can be further improved by incorporating it into other augmentation strategies, which can make MeshCut a promising baseline strategy for future data augmentation algorithms.

Highlights

Convolutional neural networks (CNNs) have demonstrated massive potential in the field of computer vision [1,2,3,4,5,6,7,8,9,10,11,12]
We propose a novel augmentation method named MeshCut
The learning rate was determined by a MultiStepLR scheduler with an initial value of 0.1, which was reduced by 10% at the 100th, 200th, and 250th epochs. This proposed MeshCut method in our experiment was not accompanied by any other data augmentations, e.g. photometric distortion or information dropping

Summary

Introduction

Convolutional neural networks (CNNs) have demonstrated massive potential in the field of computer vision [1,2,3,4,5,6,7,8,9,10,11,12]. The learning rate was determined by a MultiStepLR scheduler with an initial value of 0.1, which was reduced by 10% at the 100th, 200th, and 250th epochs This proposed MeshCut method in our experiment was not accompanied by any other data augmentations, e.g. photometric distortion or information dropping. In term of the baseline augmentation, we first resized the input image to 40 × 40 and randomly cropped a patch with a size of 32 × 32 The experimental results show that the proposed MeshCut strategy could improve upon the accuracies of all baseline models by a large margin and markedly surpassed all the listed information-dropping methods.

Findings

Discussion

Conclusion