Scene text detection with gradient guidance

Abstract

In the process of text detection, we frequently encounter numerous indistinct images, which can easily result in text omission and misdetection. Inspired by the SPSR model, we introduce gradient branching to guide the training of text detection models in order to address this problem. By preserving more image edge features, we expect to improve the text detection performance of fuzzy images and fuzzy regions. The experiment demonstrates that the gradient guidance-based text detection model can detect text in ambiguous images more accurately and reduce instances of missing and incorrect detection.