Data Augmentation by Pairing Samples for Images Classification

TL;DR

SamplePairing is a data augmentation method for image classification that creates new training samples by pixel-wise averaging two randomly chosen images and uses the first image's label. This simple, orthogonal technique effectively enlarges the training set to $N^2$ samples and acts as a strong regularizer, improving validation accuracy across diverse datasets, especially when data are scarce. The approach maintains ease of integration, requiring no changes to model architecture or labels, and benefits are most pronounced in limited-data regimes, suggesting applicability to domains like medical imaging. The paper also explores training schedules and image-overlay selection, revealing that pairing with training-set images yields greater gains than using external images. Overall, SamplePairing offers a practical, scalable boost to generalization in data-constrained image classification tasks.

Abstract

Data augmentation is a widely used technique in many machine learning tasks, such as image classification, to virtually enlarge the training dataset size and avoid overfitting. Traditional data augmentation techniques for image classification tasks create new samples from the original training data by, for example, flipping, distorting, adding a small amount of noise to, or cropping a patch from an original image. In this paper, we introduce a simple but surprisingly effective data augmentation technique for image classification tasks. With our technique, named SamplePairing, we synthesize a new sample from one image by overlaying another image randomly chosen from the training data (i.e., taking an average of two images for each pixel). By using two images randomly selected from the training set, we can generate $N^2$ new samples from $N$ training samples. This simple data augmentation technique significantly improved classification accuracy for all the tested datasets; for example, the top-1 error rate was reduced from 33.5% to 29.0% for the ILSVRC 2012 dataset with GoogLeNet and from 8.22% to 6.93% in the CIFAR-10 dataset. We also show that our SamplePairing technique largely improved accuracy when the number of samples in the training set was very small. Therefore, our technique is more valuable for tasks with a limited amount of training data, such as medical imaging tasks.

Figures (10)

• Figure 1: Overview of our SamplePairing data augmentation for ILSVRC dataset. For other datasets, the size of the original images is $32 \times 32$ and the size of the extracted patches is $28 \times 28$.
• Figure 2: Design of classification network used for CIFAR and SVHN.
• Figure 3: Changes in validation error rates for CIFAR-10 and ILSVRC datasets with and without our SamplePairing data augmentation.
• Figure 4: Changes in training error rates for CIFAR-10 and ILSVRC datasets with and without our SamplePairing data augmentation.
• Figure 5: Changes in validation and training losses for ILSVRC datasets with and without our SamplePairing data augmentation.