E-TEXTLOC: efficient text localization and extraction in real-world video scenes

Abstract

Localization of an accurate text is quite a complicated issue, especially when attempting to extract from a complex video. It is mainly due to the limitation of resources, dynamic background, and text variability. There is various artificial intelligence based methods adopting machine learning for addressing such issues encounter lower positional accuracy and incur maximized computational cost. Therefore, the proposed system introduces efficient text localization and extraction for comprehensive positional accuracy in complex videos (E-TEXTLOC). Different from conventional approaches, E-TEXTLOC facilitates sampling of video frames while MobileNetV2 is deployed towards faster localization of text. The outcome of recognized text is further refined by a verifier module, which provides a self-supervised response. Assessed on the YouTube video dataset, the proposed model accomplishes 98.2% accuracy with 29.6 ms towards generating analytical outcomes. It means the proposed model accomplishes 6-10% accuracy enhancement with a 40-50% reduction of speed in contrast to the existing system. The implications of the proposed study can be stated towards surveillance system, autonomous vehicles, and assistive devices that works in real-time.