An Enhanced Deep Learning Method to Generate Synthetic Images with Features That are Comparable to Original Images Using Neural Style Transfer

Image-to-sequence tasks have evolved greatly as a result of deep learning architectures achieving SOTA results in open-source datasets. From an industrial perspective, using these pre-trained models to attain such results mostly falls short. The main reasons are the data-hungry architecture and the limited amount of available data that is used for retraining these models. In this study, we discuss a deep learning method that generates synthetic data with properties similar to the real data in order to improve Image-to-text-based tasks. We provide a novel deep learning approach with two distinguishing steps. (1) Take an image from an OCR-based source dataset and extract features to create a new image with no text in it. (2) Use this newly formed collection of synthetic images which is very similar to the original images for fine-tuning.