Bd-Craft: Using Blind Deconvolution to Improve Scene Text Detection of the Craft Algorithm
Date of Award
12-1-2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Computer Science
First Advisor
Proceso L. Fernandez, Jr., PhD
Abstract
Recent deep learning models have demonstrated strong capabilities of text detection in scene images. Despite significant efforts to improve text recognition performance, text detection remains a difficult task, as evident in the series of Robust Reading Competitions. One of the known techniques yielding high performance for such tasks is the Character Region Awareness for Text Detection (CRAFT). Considering the good performance of CRAFT, this study explores how blind deconvolution can be used as an image pre-processing step to further improve the text detection performance of CRAFT using the ICDAR 2013 dataset in the experiments. The images in the dataset are first automatically classified into two – blurry and non- blurry images -- using a Laplacian operator. Images classified as blurry are further pre- processed using sharpening, blind deconvolution, and/or a combination of the two techniques in order to decrease the blur before running the CRAFT algorithm. Results showed that the use of blind deconvolution without sharpening is able to enhance the text detection performance of CRAFT, yielding an h-mean of 94.47% to improve the original CRAFT h-mean of 91.42% and even outperform the state-of-the-art SenseTime whose h-mean is 93.62%.
Recommended Citation
Albarillo, Clarisa V., (2022). Bd-Craft: Using Blind Deconvolution to Improve Scene Text Detection of the Craft Algorithm. Archīum.ATENEO.
https://archium.ateneo.edu/theses-dissertations/967
