Detection of texts in natural images

TL;DR

This paper presents a novel framework for detecting texts in natural images using connected components, probabilistic edge graphs, and Gabor feature-based SVM classification, achieving superior accuracy and scalability.

Contribution

The paper introduces a new method combining probabilistic edge graph analysis and Gabor SVM classification for improved text detection in natural images.

Findings

Recall and precision of 0.72 and 0.88 on ICDAR 10 dataset

Significantly better performance on natural images compared to existing methods

Scales almost linearly with high-resolution, cluttered images

Abstract

A framework that makes use of Connected components and supervised Support machine to recognise texts is proposed. The image is preprocessed and and edge graph is calculated using a probabilistic framework to compensate for photometric noise. Connected components over the resultant image is calculated, which is bounded and then pruned using geometric constraints. Finally a Gabor Feature based SVM is used to classify the presence of text in the candidates. The proposed method was tested with ICDAR 10 dataset and few other images available on the internet. It resulted in a recall and precision metric of 0.72 and 0.88 comfortably better than the benchmark Eiphstein's algorithm. The proposed method recorded a 0.70 and 0.74 in natural images which is significantly better than current methods on natural images. The proposed method also scales almost linearly for high resolution, cluttered images.