Cost-Effective and Smart Text Sensing and Spotting in Blurry Scene Images Using Deep Networks

Abstract

An intelligent transportation system facilitates smart services and applications that can revolutionize the traffic and travel experience. Driver assistance system is a crucial part of such a system that helps to improve the safety and security of passengers by mitigating on-road collisions and potential hazards. The precise sensing (localization) and spotting of scene texts and traffic signs are important for achieving higher performance in real-time. It is however affected by motion blur and camera shake noise, which makes the process of spotting complex. In this paper, we propose a robust text spotter, denoted by Blurred TextSpotter, for efficient and cost-effective spotting in blurry scene images. We address different noises, like a motion blur, Gaussian blur, camera shake noise, and inter-class interference. We apply a multi-scale contextual information enriched encoder-decoder based backbone network followed by a spatial and channel-wise attentions. We predict text masks and accurately classify words using a hardware-efficient recognition module. The experimental results on five publicly available benchmark datasets show the efficiency of the proposed text spotter in terms of detection, recognition, and spotting of curve text instances in scene images.