Natural Scene Text Detection using Deep Neural Networks

The text imparts a high level of information quickly and concisely. Therefore, retrieval of this information plays a vital role in learning and understanding for humans as it has various applications in computer vision. Various methodologies for text extraction from documents and natural scene text have been proposed in the past few decades. Two of these architectures, namely, Faster R-CNN and Effective and Accurate Scene Text detector, have been executed, analyzed, and assessed by the authors. The aim was to compare them based on their nature of pipelines, number of layers, speed of execution, and extent of accuracy. While both the networks are not perfect, the EAST detector showed exceptional results compared to other architectures. Due to a high receptive field and deeper network of DenseNet201, the inclusion of dice loss function, the model showed almost 78% accuracy.

Automated summary

The text discusses different methodologies for text extraction, focusing on comparisons between two architectures. The EAST detector showed exceptional results due to its architecture and achieved an accuracy of around 78%.