Medical Image Character Recognition Using Attention-Based Siamese Networks for Visually Similar Characters with Low Resolution

The emergence of optical character recognition (OCR) has been adopted in many domains to automate various tasks. Still, recognising visually similar characters (VSC) remains a challenging problem in the general OCR domain. Applying conventional class probability predictions by deep learning techniques may be difficult due to the limited datasets in some domains, such as medical imaging modalities. VSC recognition becomes more complicated due to the image’s low resolution and background interference. With advancements in computing power and numerical methods, techniques such as the few-shot method have been proposed to tackle the limited sample problems in training deep learning models. Still, very little work has been done regarding designing an OCR solution to deal with tiny burnt-textual data on low-resolution images with background interference while training on small samples per class. In this study, we propose an Attention-based Siamese Network to accurately recognise VSC by efficiently learning the semantic similarities between the extracted embeddings from the input images. The learned similarities and attention-focused feature extraction layer enable the proposed model to discriminate between different character classes efficiently, with only small samples available. Bayesian optimisation is used to determine optimal network parameters. We aim to set a benchmark for the performance of the Siamese network in OCR in medical image character recognition in terms of parameter size and accuracy at a determined sample size.