On automated workflow for fine-tuning deepneural network models for table detection in document images

Nowadays methods and software for extracting tables from document images and portable documents (PDF) continue to be actively developed. One of the promising approaches to this task is the usage of fine-tuned object detection models. However, this approach involves many manipulations with data preparation and training process configuration. This paper proposes an automated workflow for fine-tuning deep neural network models for the table detection in document images. It enables us to automate two sub-tasks: (i) preparing a training dataset in the PascalVOC format with image transformation and augmentation; (ii) training a table detection model by using the well-known Faster R-CNN architecture. Implementation of the workflow design simplifies the use of the approach proposed by decreasing the number of required manipulations.