Data Augmentation for Image Classification using Generative AI

Abstract

• Automated Generative Data Augmentation (AGA) : LLMs + Diffusion + Segment
• Foreground authenticity + Background diversity!!
• Object extract → Prompt + Affine.
• 15.6%, 23.5%, 64.3% improvements. Image Net, CUB, iWildCam

Comparison

• Text to image : Changed subject, Image-to-image : Details are lost. Inpainting : Corrupt

• Grad cam result will be improves as accuracy improves.

Knowledge

• SAM : SOTA segmentation model. Use YOLO for guide.

Related

• Training classifier with Gen AI image net clone | https://arxiv.org/pdf/2212.08420
• Without fine tuning (Prompt base subject) | https://arxiv.org/pdf/2305.16289 | https://arxiv.org/pdf/2312.02253