The Effectiveness of Data Augmentation in Image Classification using Deep Learning
Citations: 1,248 | Paper Published: 13 Dec 2017 | Authors: Luis Perez, Jason Wang
The paper explores different techniques for data augmentation in image classification. It compares traditional methods like cropping, rotating and flipping with GANs and proposes a new method called neural augmentation to improve the classifier's performance on various datasets.
Contributions
The main contribution of this paper is to explore the problem of data augmentation for image classification and evaluate different techniques. The authors propose a method that takes a small, curated corpus of structured data and augments it in ways that increase model performance. They compare traditional methods like cropping, rotating, flipping with GANs and introduce neural augmentation as a new technique to improve classifier's accuracy on various datasets including tiny-imagenet-200 data and MNIST.
Practical Implications
The practical implications of this paper are that it provides insights into the effectiveness of different data augmentation techniques for image classification tasks. The results suggest that traditional methods like cropping, rotating and flipping can be very effective alone but other techniques enabled by CycleGAN and similar networks show promise as well. Additionally, the proposed neural augmentation method allows a neural net to learn augmentations that best improve its ability to correctly classify images which could lead to better performance on various datasets in real-world applications.
Methods
The first approach is generating augmented data before training the classifier using techniques like cropping, rotating and flipping input images as well as GANs to generate new styles of images.
The second approach involves learning augmentations through a prepended neural net that allows it to learn which transformations best improve its ability to classify images accurately on various datasets including tiny-imagenet-200 and MNIST.
Data
The paper uses a small subset of the ImageNet dataset and three sets of images with two classes each. The first data set is taken from tiny-imagenet-200, while the other datasets are not specified in detail. A portion of all datasets was held aside for testing purposes, while the remaining images were divided by an 80:20 split between training and validation.
Results
For binary classification problem involving dogs vs cats, traditional methods like cropping, rotating and flipping input images achieved an accuracy of 85% while other advanced techniques enabled by CycleGAN showed promise as well.
On tiny-imagenet-200 dataset with 200 classes, neural augmentation method allowed learning augmentations that best improve classifier's accuracy across all test sets achieving a top-1 validation error rate reduction from baseline to around -3%.
Limitations
The study only focuses on image classification tasks and does not explore the effectiveness of data augmentation techniques for other computer vision problems such as object detection or semantic segmentation.
While neural augmentation shows promise as an approach for learning augmentations that best improve classifiers' accuracy across different datasets but its computational cost may be prohibitive when dealing with large-scale applications like autonomous driving systems or medical imaging diagnosis tools