A survey on Image Data Augmentation for Deep Learning

A survey on Image Data Augmentation for Deep Learning Connor Shorten Taghi M. Khoshgoftaar

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

A survey on Image Data Augmentation for Deep Learning 2019 Journal Of Big Data 10.1186/s40537-019-0197-0 shorten2019-a-survey-on-image-data-augmentatio shorten2019-a-survey-on-image-data-augmentatio.txt, the project's staged copy 5cb5b602be8f94d70c264f5d764012b93694847b cc-by machine-learning

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them image Data Augmentation techniques, from geometric transforms to GANs, as a remedy for overfitting on limited data Quotation withheld because this record is not cleared for redistribution. what to do about overfitting when the application domain has no big data to be had Quotation withheld because this record is not cleared for redistribution. a taxonomy of image augmentation split into data warping and oversampling, together with the statement that the split is not exclusive Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. overfitting - a model learning a function of very high variance that models its training data perfectly Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. mechanisms AUGMENTATION IS AN ASSUMPTION, AND THE PAPER STATES IT: that more information can be extracted from the dataset already held. Everything else follows from whether that is true of a given dataset Quotation withheld because this record is not cleared for redistribution. mechanisms WHAT AUGMENTATION IS ACTUALLY TEACHING: the invariances a recogniser must survive - viewpoint, lighting, occlusion, background, scale - are baked into the DATA rather than into the architecture Quotation withheld because this record is not cleared for redistribution. mechanisms SIMULATING BIG DATA RATHER THAN OBTAINING IT: overfitting is not much of a problem when data is abundant, so augmentation gives a limited dataset the characteristics of an abundant one Quotation withheld because this record is not cleared for redistribution. an unsafe transformation could be made useful rather than harmful by softening the label, so the model learns uncertainty instead of a falsehood - but building those labels is too costly Quotation withheld because this record is not cleared for redistribution. oversampling works by re-balancing the class distribution so the model is not biased toward answering with the majority class Quotation withheld because this record is not cleared for redistribution. medical imaging is short of data FOR STRUCTURAL REASONS, and the paper names rarity first: `Quotation withheld`. The remaining three reasons it lists - patient privacy, the need for expert labellers, and the cost of the scans - are separated from this clause by a journal running header in the staged text and are not inside this span. Quotation withheld because this record is not cleared for redistribution. rotation safety is set by one parameter: small rotations help digit recognition and large ones destroy the label, so the same transform is safe and unsafe at different magnitudes Quotation withheld because this record is not cleared for redistribution. cropping and translation differ in what they preserve - cropping shrinks the input, translation keeps the spatial dimensions and pads the vacated space Quotation withheld because this record is not cleared for redistribution. colour transformations trade colour bias for spatial characteristics, which is a gain only where colour is not the distinguishing feature Quotation withheld because this record is not cleared for redistribution. a kernel filter can be rebuilt as a convolution layer with the right padding and activation, so the augmentation duplicates what the network already does Quotation withheld because this record is not cleared for redistribution. AUGMENTATION CANNOT CREATE A CLASS THAT IS ABSENT: no method surveyed will produce a golden retriever from a dataset of bulldogs. It can lessen biases of lighting, occlusion, scale and background, and it cannot supply missing content Quotation withheld because this record is not cleared for redistribution. The stated bias limit in the paper's own words, which bounds every technique it surveys Quotation withheld because this record is not cleared for redistribution. The survey covers IMAGE data only, with transfer to other data domains asserted rather than shown Quotation withheld because this record is not cleared for redistribution. The taxonomy the survey is organised by does not partition its own subject - warping and oversampling are stated not to be mutually exclusive Quotation withheld because this record is not cleared for redistribution. safety is domain-dependent, which is precisely what makes a generalisable augmentation policy hard Quotation withheld because this record is not cleared for redistribution. there is NO image processing function that cannot flip a label at some distortion magnitude - the strongest form of the claim, stated by the authors Quotation withheld because this record is not cleared for redistribution. constructing refined labels for every unsafe augmentation is computationally expensive, which is why the cheap option is to avoid unsafe transforms Quotation withheld because this record is not cleared for redistribution. many experiments deliberately restrict themselves to a subset of a large dataset in order to SIMULATE a limited-data problem rather than face a real one Quotation withheld because this record is not cleared for redistribution. kernel filters are very close to the internal mechanism of a CNN, which is the argument for making them a layer rather than a dataset operation Quotation withheld because this record is not cleared for redistribution. in medical imaging the gap between training and test data is more complex than position and translation, so geometric transformations have limited scope there Quotation withheld because this record is not cleared for redistribution. colour transformations can discard the information the task depends on - darkening may hide the object entirely Quotation withheld because this record is not cleared for redistribution. transformations that may alter the label, such as translation or random cropping, must be checked by hand Quotation withheld because this record is not cleared for redistribution. SMOTE and Borderline-SMOTE were built to interpolate new points for imbalanced TABULAR data and are carried over to images Quotation withheld because this record is not cleared for redistribution. transfer learning is explained by the anecdote that a pianist learns guitar faster, and augmentation by analogy with imagination and dreaming Quotation withheld because this record is not cleared for redistribution. Data Augmentation Data Augmentation Quotation withheld because this record is not cleared for redistribution. Generalizability Generalizability Quotation withheld because this record is not cleared for redistribution. overfitting overfitting Quotation withheld because this record is not cleared for redistribution. data warping data warping Quotation withheld because this record is not cleared for redistribution. oversampling oversampling Quotation withheld because this record is not cleared for redistribution. Dropout Dropout Quotation withheld because this record is not cleared for redistribution. Batch normalization Batch normalization Quotation withheld because this record is not cleared for redistribution. Transfer Learning Transfer Learning Quotation withheld because this record is not cleared for redistribution. Pretraining Pretraining Quotation withheld because this record is not cleared for redistribution. translational invariances translational invariances Quotation withheld because this record is not cleared for redistribution. training dataset training dataset Quotation withheld because this record is not cleared for redistribution. regularization technique regularization technique Quotation withheld because this record is not cleared for redistribution. big data big data Quotation withheld because this record is not cleared for redistribution. GAN samples GAN samples Quotation withheld because this record is not cleared for redistribution. safety safety Quotation withheld because this record is not cleared for redistribution. class imbalance class imbalance Quotation withheld because this record is not cleared for redistribution. Random Oversampling Random Oversampling Quotation withheld because this record is not cleared for redistribution. SMOTE SMOTE Quotation withheld because this record is not cleared for redistribution. refined labels refined labels Quotation withheld because this record is not cleared for redistribution. Neural Style Transfer Neural Style Transfer Quotation withheld because this record is not cleared for redistribution. positional bias positional bias Quotation withheld because this record is not cleared for redistribution. kernel filters kernel filters Quotation withheld because this record is not cleared for redistribution. PatchShuffle PatchShuffle Quotation withheld because this record is not cleared for redistribution. color space transformations color space transformations Quotation withheld because this record is not cleared for redistribution. noise injection noise injection Quotation withheld because this record is not cleared for redistribution. Data Augmentation - A DATA-SPACE SOLUTION, defined by WHERE it intervenes rather than by what it does: every other remedy for overfitting changes the model, and this one changes the dataset Quotation withheld because this record is not cleared for redistribution. data warping - the label is PRESERVED and an existing image is transformed Quotation withheld because this record is not cleared for redistribution. oversampling - a NEW instance is synthesised and added. The paper draws the distinction and then says in the next breath that it does not partition: GAN samples can be stacked with random cropping Quotation withheld because this record is not cleared for redistribution. data warping —distinct-from→ oversampling data warping oversampling holds when: the two are NOT mutually exclusive - GAN samples can be stacked with random cropping, so a single augmentation can be both · strength of the relation: 0.8 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says the two augmentations do NOT form a mutually exclusive dichotomy; distinct-from is hedged. Quotation withheld because this record is not cleared for redistribution. Transfer Learning —distinct-from→ Pretraining Transfer Learning Pretraining strength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but endpoint by anaphora: 'This differs from Transfer Learning' - 'This' = Pretraining (previous sentences). Quotation withheld because this record is not cleared for redistribution. Dropout —is-a→ regularization technique Dropout regularization technique strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. Batch normalization —is-a→ regularization technique Batch normalization regularization technique strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. Data Augmentation —causes→ overfitting Data Augmentation overfitting direction on the measure: reduces · the relation itself is separately evidenced · strength of the relation: 1 · direction on the measure (stated by the tagger, no quotation): reduces · the relation stated: augmentation gives a limited dataset the characteristics of big data, and overfitting is not much of an issue when data is abundant Quotation withheld because this record is not cleared for redistribution. AlexNet's experiments used augmentation to enlarge the dataset by a factor of 2048. Quotation withheld because this record is not cleared for redistribution. GAN-based augmentation beat classic augmentation on liver lesion classification on both sensitivity and specificity Quotation withheld because this record is not cleared for redistribution. Horizontal flipping is much more common than vertical flipping. Quotation withheld because this record is not cleared for redistribution. converting to grayscale costs about three points of classification accuracy against RGB on ImageNet and PASCAL VOC Quotation withheld because this record is not cleared for redistribution. in a head-to-head comparison of geometric against photometric augmentation, cropping produced the most accurate classifier Quotation withheld because this record is not cleared for redistribution. randomly swapping pixels inside a small window lowered CIFAR-10 error against the same architecture without it Quotation withheld because this record is not cleared for redistribution. Blurring training images could make models more resistant to motion blur at test time. Quotation withheld because this record is not cleared for redistribution. the safety of an augmentation defined as label preservation, with the cat/dog against 6/9 example safety Quotation withheld because this record is not cleared for redistribution. the factor by which AlexNet's augmentation multiplied the dataset, and the error reduction claimed for it Quotation withheld because this record is not cleared for redistribution. sensitivity and specificity before and after GAN-based augmentation on liver lesion classification Quotation withheld because this record is not cleared for redistribution. the CIFAR-10 error rate with and without PatchShuffle regularization, and the hyperparameters that achieved it Quotation withheld because this record is not cleared for redistribution. the accuracy drop from grayscale against RGB Quotation withheld because this record is not cleared for redistribution. the rotation degree parameter, on which the safety of a rotation augmentation depends rotation degree Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1186/s40537-019-0197-0.