This study analyzed how the performance of various neural network architectures responds to offline image augmentation using standard geometric and photometric transformations. We evaluated a composite training-data enrichment protocol using two multiclass color image classification tasks: a dataset of leg photographs classified according to the CEAP system (sourced from open-source repositories; 20,971 images, seven classes) and a dermatological image dataset (6786 images, eight classes). To enrich the initially balanced subsets, transformed images were added to classes underrepresented in the original datasets, while additional real images were added to the remaining classes. Twelve architectures were investigated, including classical CNNs, vision transformers, and a modernized CNN. Across the 12 architectures, 11 augmentation levels, and 20 runs per configuration, mean accuracy ranged from 0.55 to 0.68 (leg photographs) and from 0.59 to 0.79 (dermatological images). The expansion levels yielding the highest observed mean accuracy varied across configurations, with descriptive maxima occurring at levels from 0% to 100%. For each model–dataset configuration, the highest mean accuracy across the eleven levels was treated as a descriptive maximum. Standard deviations across runs characterized the variability of the training process, whereas standard errors reflected the uncertainty of mean estimates given a fixed test set. Transformer-based configurations demonstrated lower average run-to-run standard deviations; however, this tendency was configuration- and dataset-dependent, warranting further research. An exploratory augmentation-as-biased-variance-reduction (ABVR) model is proposed to describe the observed accuracy curves. Its fitted coefficients and maximizers summarize the available data and have not been independently validated.

This study examines the impact of data augmentation on machine learning performance, focusing on how synthetic data influences various neural network architectures. Common issues such as limited data, class imbalance, and poor coverage often lead to low model metrics, and data augmentation is frequently used to address these problems. The research aims to identify the optimal proportion of synthetic data, assess its effects across different architectures, and analyse the impact of augmenting only the rare classes in two multi-class colour-image tasks: CEAP-graded leg photographs collected from public web sources (20,971 images, 7 classes) and an open dermatological dataset from a different domain, on which the whole protocol was replicated independently (6,786 images, 8 classes). Twelve architectures were evaluated, including convolutional, transformer-based, and hybrid models. Across 12 architectures, 11 augmentation levels and 20 runs per configuration, mean accuracy varied within 0.55-0.68 (leg photographs) and 0.59-0.79 (dermatological images), and no single augmentation ratio was optimal for all models: the best ratio ranged from 0% (Inception) to 100% (ViT, DenseNet161). The differences between the best and the baseline ratio are small (0.01-0.03) but, referred to the standard error of the mean over 20 runs (0.002-0.009) rather than to the run-to-run standard deviation, they remain significant after Holm correction for 10 of 12 models on the leg photographs and for 4 of 12 on the dermatological images. Transformer-based models showed the lowest run-to-run dispersion of accuracy, while convolutional networks showed the highest. To explain these observations we propose a model of augmentation as biased variance reduction, in which the effect of the synthetic share is the sum of a variance term decreasing as 1/(1 + α) and a label-shift bias growing as the square of the share; the model derives the model-specific optimum, bounds the attainable gain by a quarter of the run-to-run variance, and reproduces the contrast between the two datasets with a single fitted coefficient that differs between them by a factor of eighteen.