Synthetic Data Generation Pipeline for Private ID Cards Detection
Diksha Bothra, Sanket Dixit, Debi Prasanna Mohanty, Mohammad Haseeb, Shyamji Tiwari, Amit Tulsidas Chaulwar · 2023
Modern deep learning methods have achieved remarkable performance on many real-world problems. However, the availability of large real data is a prerequisite for these methods to work. For applications like private ID card detection, it is difficult to gather real-world data because of privacy reasons and the sensitive nature of the data. Therefore, we propose a pipeline for synthetic card generations with random details followed by different methodologies for inserting them in random images such that they present different real-world scenarios such as cards held in the hand, placed on flat surfaces, etc. To simulate the environmental conditions affecting the card appearance in the image, we also propose to apply different image transformations to cards like rotation, blurring, etc. Finally, we also show the quality of the generated dataset by training the lightweight EfficientDet-Lite1 model with it and then testing it with images containing the real ID cards.