A MULTI-HYBRID FRAMEWORK FOR IMAGE DATA AUGMENTATION IN A RANDOMIZED SETTING

Abstract

Image Data Augmentation (IDA) is fundamental for improving model robustness and reducing overfitting; however, many existing approaches are tightly integrated in specific deep learning models, limiting their applicability primarily to classification tasks. This study proposes a model-agnostic, stand-alone multi-hybrid IDA framework that operates independently of any machine learning (ML) model. The framework is implemented as a Python-based system integrating five common transformation techniques: rotation, translation, zooming, flipping, and color-space manipulation. The proposed Randomized Framework was benchmarked against three established augmentation pipelines with a Sequential mode serving as an internal baseline. Performance evaluation employed dataset diversity, class preservation, execution time, and classifier validation using K-Nearest Neighbour (KNN) and Support Vector Machine (SVM) models on features extracted via ResNet-50. Results demonstrate that the Randomized Framework achieved the highest diversity score (0.3767) with optimal execution time (5.20 s), alongside superior generalization evidenced by the lowest overfitting gap (0.03) and strong classification performance (AUC = 0.9596 for KNN). Classification accuracy improved from 0.91–0.93 to 0.95–1.00 after augmentation. The study concludes that randomized multi-hybrid augmentation offers an effective balance between diversity and efficiency and, due to its specific machine learning model- independent design, it is well suited for diverse computer vision applications. Future work will investigate additional transformations, deep learning–based augmentation pipelines, and scalable batch-level optimization.