Applicable Scene Text Detection Based on Semantic Segmentation

Abstract

Abstract In the past few years, scene text detection problem has attracted much attention and widely studied for many applications. Segmentation based methods are quite popular in this filed, it was not only for the simple post-processing procedure, but also for the ability of handle the various shape of text. In this manuscript, a method named CFPM-IDB was proposed which based on a new decoder module Consolidated Feature Pyramid Module (CFPM) and a light weight segmentation head Improved Differentiable Binarization (IDB). The CFPM module originated from Feature Pyramid Networks and achieved an idea trade-off between local details and macro semantic information. IDB module performed binarization process by threshold map and probability map. Combining them results in CFPM-IDB, which improved the performance of scene text detection, both precision and efficiency. Experimental results on two different datasets proved the promising advance in the accuracy and speed of this model.