Bridging Reality and Synthetics: Optimizing Image Classification with Hybrid AI-Generated and Real-World Datasets

Abstract

The rapidly growing revolution of generative Artificial Intelligence software has moved into the counseling and disseminating synthetic images, thereby establishing a new paradigm for machine learning models. This study investigates the impact of combining real-world and AI-generated synthetic images on the performance of image classification models. Using three traffic-related datasets—potholes, speed bumps, and traffic lights—we applied data augmentation and tested seven configurations with varying real-to-synthetic image ratios. The DenseNet201 model, fine-tuned with the Adam optimizer, was used for all experiments. Results show that a 1:3 real-to-synthetic ratio enhances classification accuracy and generalization, with the highest validation accuracy reaching 97.36%. Our findings demonstrate that synthetic data, when properly integrated, serves as a cost-effective and scalable complement to real data, especially in scenarios with limited labeled samples.