Volume 14, Number 1

Investigating the Effect of BD-CRAFT to Text Detection Algorithms

  Authors

Clarisa V. Albarillo¹ and Proceso L. Fernandez, Jr.², ¹Don Mariano Marcos Memorial State University, Philippines, ²Ateneo de Manila University, Philippines

  Abstract

With the rise and development of deep learning, computer vision and document analysis has influenced the area of text detection. Despite significant efforts in improving text detection performance, it remains to be challenging, as evident by the series of the Robust Reading Competitions. This study investigates the impact of employing BD-CRAFT – a variant of CRAFT that involves automatic image classification utilizing a Laplacian operator and further preprocess the classified blurry images using blind deconvolution to the top-ranked algorithms, SenseTime and TextFuseNet. Results revealed that the proposed method significantly enhanced the detection performances of the said algorithms. TextFuseNet + BD-CRAFT achieved an outstanding h-mean result of 93.55% and shows an impressive improvement of over 4% increase to its precision yielding 95.71% while SenseTime + BD-CRAFT placed first with a very remarkable 95.22% h-mean and exhibited a huge precision improvement of over 4%.

  Keywords

Blind Deconvolution, Computer Vision, Image Classification, Information Retrieval, Image Processing.