SDINet Scheme for Generalized Text Detection in Scene and Document Images

Abstract

Text Detection is an essential intermediate step in optical character recognition (OCR). OCR applied in scene text images is helpful for applications such as traffic signs and vehicle number plate recognition. OCR applied in the document text images help digitise and analyse the documents. Hence, a robust text detection system is needed to detect the text exceptionally well given an arbitrary text image. In this work, we address text detection in images using our Scene or Document Image Network (SDINet). During the training of the model, a Weighted Loss (WL) is designed to better update the training parameters according to the input image type. A classification model is designed that helps us to find the WL By classifying an input image as a scene text type image or document text type image. The novelty of our approach is in the fact that the training parameters of the model are updated according to the input image type. Our approach shows comparative results in all the evaluation parameters for scene text and document text datasets. Specifically, when compared to PSENet, experimental results show that our SDINet approach improves the recall by more than 1%, and F-score is increased by approximately 1% for SCUT-CTW 1500 dataset. [30].