An Efficient Scene Text Spotter and Detector Using Deepnet

Abstract

Scene text detection is a complex and challenging task due to various environmental factors, such as varying illuminations, lighting conditions, tiny and curved texts, and more. Many existing works in scene text detection focus on improving model accuracy but often overlook the need for efficiency, resulting in heavy-weight models that require significant processing resources. This paper introduces a novel efficient model to address the objectives of improving accuracy and efficiency in scene text detection.This paper proposes a new hybrid model for text detection in images that uses ResNet50 with AtrousSpatial Pyramid Pooling (ASPP)  based on Efficient and Accurate Scene Text (EAST) algorithm with A technique for suppressing duplicate text detections that is more lenient than traditional non-maximum suppression.The proposed method is designed to improve the efficiency and accuracy of text detection, Experiments on the IIIT-ILST, ICDAR2015 and ICDAR2019, MSRA-TD500 dataset, show that the proposed method achieves state-of-the-art results