Advanced Random Mix Augmentation : Data Augmentation Method for Improving Performance of Image Classification in Deep Learning Using Unduplicated Image Processing Combinations

Jonghoon Im, Jun Younes Louhi Kasahara, Hiroshi Maruyama, Hajime Asama, Atsushi YamashitaJournal of the Japan Society for Precision Engineering 2023 · DOI

Abstract

Data augmentation is a commonly used method for improving deep learning models in image classification. By adding slightly modified images that do not change the label of the original image to the training data set, the trained model becomes more robust against diverse characteristics of the input image. In this study, we propose a new data augmentation method by improving a previously-known random augmentation method. Our method consists of three steps; 1) determine the set of image modification operators and the number of augmented images, 2) determine the sequence of the image modification operators so that no duplicated sequences are generated, and 3) apply the sequence to augment images. The variety of augmentation is further increased by randomly determining the level (intensity) and the weight of combining the sequences. We applied our method on the CIFAR dataset and show that our method outperforms existing methods.

Introduction: The Challenge of Redundancy in Data Augmentation

Deep learning models are notorious for overfitting—performing exceptionally well on training data but failing to generalize to new, unseen images. Data augmentation (DA) is the standard remedy, artificially expanding the training set by applying transformations like rotations or color shifts.

However, popular modern DA methods like AugMix or Random Mix Augmentation (RMA) rely on purely random selection. This randomness often leads to a "collision" problem: the same combinations of operations are applied multiple times, while other potentially useful combinations are never explored. This redundancy limits the diversity of the training data. The authors sought to create a method that guarantees a wider variety of unique augmented images to better "fill" the data space.

Advanced Random Mix Augmentation (ARMA)

To address this, we introduce Advanced Random Mix Augmentation (ARMA). The core innovation of ARMA is a systematic sampling approach that guarantees every augmented image within a training cycle is generated from a unique sequence of processing operations.

The ARMA Framework

The ARMA process involves four primary stages:

  1. Combination Size Determination: Defining the number of augmented versions per image and the depth of operations.
  2. Unique Combination Selection: Instead of per-image random selection, ARMA pre-calculates all unique operation combinations and samples them without replacement. This ensures zero redundancy across the augmented set. For 5 million generated images, existing methods (RMA) produced over 116,000 redundant combinations, while ARMA produced zero.
  3. Augmentation Application: Operations (e.g., Rotate, Solarize, Translate) are applied with a randomly assigned intensity "level" (1–7).
  4. Multi-Path Mixing: Multiple uniquely transformed paths are blended together and then mixed with the original image using weights derived from a Dirichlet distribution.
The ARMA pipeline illustrating the transformation of an input image through multiple unique operation paths and their subsequent mixing.

Experimental Evaluation

We evaluated ARMA using standard benchmarks, including CIFAR-10 and CIFAR-100, as well as their "corrupted" counterparts (CIFAR-10-C and CIFAR-100-C) which test model robustness against 15 types of real-world distortions like noise, blur, and weather effects.

Key Results

ARMA consistently outperformed state-of-the-art methods across various metrics:

  • Superior Accuracy: On CIFAR-100, ARMA achieved an error rate of 22.15%, outperforming AugMix (23.30%) and RMA (22.52%).
  • Enhanced Robustness: The method showed significant gains on corrupted datasets, reducing the error rate on CIFAR-100-C to 34.96%, compared to 35.76% for AugMix.
  • Architecture Agnostic: The performance benefits were consistent across different model architectures, including standard CNNs, DenseNet, and WideResNet, showing that the benefits of unique augmentation combinations are universal.
Comparison of error rates across CIFAR datasets. ARMA (Proposed) consistently achieves the lowest error rates compared to standard and existing augmentation methods.

Insights and Conclusion

Our findings highlight that diversity is more critical than quantity in data augmentation. By eliminating redundant operation sequences, ARMA effectively "fills" the data space more efficiently than purely random methods. Furthermore, we identified a "sweet spot" for transformation intensity (Level 5), beyond which images often lose their semantic meaning and become "unlearnable."

Future work will focus on optimizing the selection of specific image processing operations rather than treating all operations as equally likely, and scaling this unique combination logic to larger, more complex computer vision tasks beyond simple classification.

Summary updated

Abstract

Data augmentation is a commonly used method for improving deep learning models in image classification. By adding slightly modified images that do not change the label of the original image to the training data set, the trained model becomes more robust against diverse characteristics of the input image. In this study, we propose a new data augmentation method by improving a previously-known random augmentation method. Our method consists of three steps; 1) determine the set of image modification operators and the number of augmented images, 2) determine the sequence of the image modification operators so that no duplicated sequences are generated, and 3) apply the sequence to augment images. The variety of augmentation is further increased by randomly determining the level (intensity) and the weight of combining the sequences. We applied our method on the CIFAR dataset and show that our method outperforms existing methods.
BibTeX
@article{im2023advanced,
  author = {Jonghoon Im and Jun Younes Louhi Kasahara and Hiroshi Maruyama and Hajime Asama and Atsushi Yamashita},
  title = {{Advanced Random Mix Augmentation : Data Augmentation Method for Improving Performance of Image Classification in Deep Learning Using Unduplicated Image Processing Combinations}},
  journal = {Journal of the Japan Society for Precision Engineering},
  year = {2023},
  doi = {10.2493/jjspe.89.105},
}
Figure