Abstract:Data augmentation has been proved to be a critical and effective way to alleviate over-fitting of deep learning model. Region-level removal is one of state-of-the-art solutions, which can not only synthesize vicinity samples, but also improve generalization of model. However, region removing using random strategy tends to make the training samples suffer from excessive information loss and the introduction of negative noise. In this paper, we propose a novel data augmentation scheme called Self-Replacement-and-Keeping-Augment (SRK-Augment), which exploits self-deformation data as the replacement template and keeps discriminative parts guided by Class Activation Map (CAM) in input image. Concretely, we firstly exploit Grad-CAM++ algorithm to calculate the CAM mask of the input image, and design a patch-shuffling mechanism ( PS-operator ) to obtain the structural self-deformation template. Then, we utilize the self-deformation template to fill the information removal area, as well as we apply the binary CAM mask to recover the discriminative regions. Finally, these augmented data will be randomly used for model training. The proposed method is simple to implement and can be incorporated with existing augmentation strategies with low computational cost. Extensive experiments are conducted on the challenging datasets. With the help of the SRK-Augment strategy, the performances of DCNNs have achieved obvious improvements. On CIFAR-10 dataset, the Top-1 error rate is dropped by 2.07% at most; On CIFAR-100 dataset, the Top-1 error rate is decreased by up to 3.73%; On Mini-ImageNet dataset, the maximum decline of the Top-1 error rate is 3.38%; On Pascal VOC dataset, the mean Average Precision increases by a maximum of 1.38%. Experimental results manifest the effectiveness and generality of the proposed method.

Interpretability-Mask: a Label-Preserving Data Augmentation Scheme for Better Classification

SRK-Augment: A Self-Replacement and Discriminative Region Keeping Augmentation Scheme for Better Classification.

AdvMask: A Sparse Adversarial Attack Based Data Augmentation Method for Image Classification

KeepAugment: A Simple Information-Preserving Data Augmentation Approach

What Are Effective Labels for Augmented Data? Improving Calibration and Robustness with AutoLabel

FenceMask: A Data Augmentation Approach for Pre-extracted Image Features

Exploring Data Augmentations on Self-/Semi-/Fully- Supervised Pre-trained Models

Data-Efficient Augmentation for Training Neural Networks

Saliency information and mosaic based data augmentation method for densely occluded object recognition

The Performance Research of the Data Augmentation Method for Image Classification

Robust Classification by Coupling Data Mollification with Label Smoothing

FakeMix Augmentation Improves Transparent Object Detection

LA3: Efficient Label-Aware AutoAugment

In-and-Out: a data augmentation technique for computer vision tasks

A Study on the Impact of Data Augmentation for Training Convolutional Neural Networks in the Presence of Noisy Labels

Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition

Understanding the Detrimental Class-level Effects of Data Augmentation

Soft Augmentation for Image Classification

Revealing the Two Sides of Data Augmentation: An Asymmetric Distillation-based Win-Win Solution for Open-Set Recognition

Stochastic Batch Augmentation with An Effective Distilled Dynamic Soft Label Regularizer

Augmentation Invariant Training