Enhancing text recognition on Tor Darknet images

Abstract

Text Spotting can be used as an approach to retrieve information found in images that cannot be obtained otherwise, by performing text detection first and then recognizing the located text. Examples of images to apply this task on can be found in Tor network images, which contain information that may not be found in plain text. When comparing both stages, the latter performs worse due to the low resolution of the cropped areas among other problems. Focusing on the recognition part of the pipeline, we study the performance of five recognition approaches, based on state-ofthe-art neural network models, standalone OCR, and OCR enhancements. We complement them using string-matching techniques with two lexicons and compare computational time on five different datasets, including Tor network images. Our final proposal achieved 39,70% precision of text recognition in a custom dataset of images taken from Tor domains.

Rectification and Super-Resolution Enhancements for Forensic Text Recognition

“…When applied to similar characters, these approaches may highlight these alterations by deblurring or enhancing key areas, which can potentially penalize similar character recognition. Such a penalty can be prevented by implementing lexicons, string-matching techniques and the average-edit distance to measure word closeness better and avoid recognition mistakes [ 23 ].…”

“…We can attribute the lower score obtained in these works [ 22 , 23 ] to different image quality factors, such as colour distribution, brightness, and partial occlusion, some of which we illustrate in Figure 2 . Other problems affecting the text recognition task include multiple fonts or languages in the same image, character similarity, lighting conditions or even mistakes when labelling the images [ 24 ].…”

“…In the first, Blanco et al [ 22 ] obtained an F-Measure of 0.57 in the text detection task, but their performance on text recognition with state-of-the-art techniques was much lower. Through the use of dictionaries and string-matching techniques, Blanco et al [ 23 ] improved the score of the recognition stage to 0.3970.…”

“…Other problems affecting the text recognition task include multiple fonts or languages in the same image, character similarity, lighting conditions or even mistakes when labelling the images [ 24 ]. Although these are frequent problems in state-of-the-art datasets, the issues of low resolution and oriented text are the most remarkable in both CSA and Tor-based images [ 23 ]. In Tor images, it is very common to fit several documents in a single picture to express quantity or product variations.…”

“…In this paper, we address the problem of performing text recognition on non-horizontal and low-resolution text [ 25 ], by enhancing images using two different techniques; rectification networks [ 26 , 27 ], which correct an image’s orientation to reduce transcription mismatches, and super-resolution techniques, which improve the image quality before recognition. We combine these two tasks on CSA focused images as well as on Tor darknet images [ 22 , 23 ] in order to retrieve information that can be of use to identify potentially illegal activities.…”