From Synthetic To Real: Enhancing Deep Learning Models With Generative Adversarial Networks For Efficient Data Utilization In Automatic Retail Stores

One of the major challenges in developing deep learning models for automatic retail stores is the availability of diverse and accurate data. Achieving high accuracy requires a large amount of data that is specifically collected and labeled for the task at hand. However, this process is time-consuming and prone to potential errors during labeling, which can affect model training. To mitigate this challenge, data augmentation methods have been widely adopted to train models when data is limited. However, the effectiveness of data augmentation is not always guaranteed, as it still relies on the original data. To address this limitation, a new approach has been developed that utilizes synthetic data generated through 3D scanning of objects. However, the challenge lies in ensuring that this synthetic data accurately represents real-world data to maintain model accuracy. In this paper, we propose an approach that integrates adversarial learning into the YOLOv8 model and trains it using synthetic data. We then apply this method to automatic retail stores to demonstrate its efficacy in real-life scenarios. Experimental results on the AIC-22 dataset reveal significant improvements compared to the original YOLOv8 version. The proposed approach achieves around 6.7% increase in precision, a 7.1% improvement in recall, and a 10.3% enhancement in mAP while maintaining the same inference speed of 127 fps on the RTX 3090Ti.