Deep Learning based Framework for Text Detection

Abstract

At present, object detection is estimated as a core assignment in the province of computer vision. Strategies for text detection dependent on deep learning have been applied in variety of fields, like smart transportation systems and self-sufficient driving system. In recent times deep learning based strategies used for localization of text have acquired cheerful results with respect to diverse datasets, but, they typically fall behind when presented to precarious circumstance. Still there is a challenge among speed and precision for text discovery. Customarily for text detection generally the two phase detectors like Faster-RCNN, Fast-RCNN and R-CNN were investigated. In this work, a deep learning based framework for detecting text in scene pictures is proposed. Proposed framework is based on one stage object detection system identified as YOLOv5 which is nearly quicker and precise, difference to existing object detectors. It gives quick and exact identification of text from basic scene pictures. Analyses performed on condition of craftsmanship datasets including ICDAR 2013, ICDAR 2015, ICDAR 2003, SVT, MSRA-TD-500 show that the proposed framework essentially performs better aside from ICDAR2015, contrast with existing strategies regarding rightness. Proposed framework obtains optimum results for ICDAR2013 dataset. The size of model produced through our framework is small.