Image Enhancement for Improved OCR and Deck Segmentation in Shipbuilding Document Images

Nickan Sadri, Jean Desir, Wael Khallouli, Mohammad Shahab Uddin, Samuel F. Kovacic, Andres Sousa-Poza, Mustafa Cannan, Jiang Li

Abstract: In this paper, we develop a CycleGAN model to enhance shipbuilding document images for downstream tasks including optical character recognition (OCR) and ship deck segmentation. We train the CycleGAN model with public image denoising datasets and utilize the trained model as an image denoising model to enhance the Navy’s Military Sealift Command (MSC) document images. We then apply our previous optical character recognition (OCR) pipeline and deck segmentation model to the enhanced MSC documents images. Experiment results show that the proposed model achieved promising improvements in signal-to-noise ratio (SNR), OCR accuracy, and deck segmentation.