Impact of Advanced Data Augmentation and Synthetic Samples on Model Generalization in Small- Scale Computer Vision Tasks

Abstract

In many specialized computer vision applications, such as smart wearable sensor monitoring and industrial quality control, acquiring large-scale labeled datasets is often prohibitive due to privacy concerns, operational costs, and the rarity of target events. Consequently, convolutional neural networks (CNNs) trained on limited domain-specific data suffer from severe overfitting and poor out-of-distribution generalization. This paper systematically investigates the effectiveness of various data augmentation strategies in enhancing CNN robustness and feature invariance under data-scarce conditions. We evaluate three distinct training frameworks: a baseline model trained on unaugmented raw inputs, a model utilizing standard geometric (rotation, flipping, cropping) and color-space transformations (jitter, brightness adjustment), and an advanced framework incorporating synthetic sample generation using CutMix combined with domain-randomized background-swapping. Experiments were conducted across high-variance sensor and industrial inspection datasets. Quantitative results reveal that traditional augmentations offer incremental gains, whereas the synthetic-aware training strategy significantly mitigates overfitting, achieving a substantial increase in classification accuracy and F1-score. Crucially, the proposed synthetic-aware approach demonstrates exceptional performance retention and structural robustness when subjected to adverse environmental conditions, including severe occlusion, lighting fluctuations, and background noise. These findings underscore the practicality of synthetic augmentation pipelines for deploying reliable vision models in data-constrained industrial and wearable environments.