Self-Conditioned Generative Data Augmentation for Image Classification

Data augmentation is a widely used technique that not only increases the size of the training dataset but also improves the generalization ability of deep learning models. Traditional image augmentation methods primarily rely on geometric transformations. More recently, approaches utilizing pre-trained generative models have been proposed to further enhance dataset diversity by synthesizing new data. However, these approaches often suffer from several limitations, including the generation of distorted objects, repetitive or unrealistic background patterns, and a dependency on carefully crafted prompts to produce semantically meaningful results. In this paper, we propose Self-Conditioned Generative Data Augmentation (SCGDA), a novel pipeline that generates images under diverse conditions while preserving the visual cues of the target objects without requiring model tuning or manual prompt engineering. SCGDA effectively generates image through three automated generative steps, reflecting diverse conditions while preserving the essential visual cues of the target object. We applied SCGDA to the ResNet models and evaluated it on the reduced ImageNet datasets. Experimental results show that our proposed method improves classification performance by up to 8.8% on the ResNet model, demonstrating its effectiveness.