1. Brief introduction
Image classification models are able to predict data that has the same distribution as the training data. However, in real-world scenarios, the input data may change. For example, when using a different camera for inference, the lighting conditions, contrast, color distortion, etc. may differ from the training set and significantly affect the performance of the model. To address this challenge, Hendrycks et al. proposed the AugMix algorithm . Can be applied to any image classification model to improve its robustness and uncertainty estimation.
AugMix is a data augmentation technique that generates augmented variations of each training image. When combined with a consistency loss, it encourages the model to make consistent predictions for all versions of the same image. Although it takes longer to train models with these augmented versions of the data, the resulting models become more stable, consistent, and resistant to a variety of inputs.
Augmented versions of training images are generated by applying three parallel chains of one to three randomly chosen augmentation operations (e.g. translation, cropping, and contrast), with randomly determined strengths. These chains are then combined with the original image with different weights to produce a single version of the augmented image. The augmented version contains multiple sources of randomness, including the choice of operations, the strength of those operations, the length of the augmented chain, and mixing weights.
For each training image, AugMix generates two augmented versions (augmix1 and augmix2) that preserve the semantic content of the image.
You can see an example of the image and its enhanced version in the image below.