Detecting Textual Information in Images from Onion Domains Using Text Spotting

Abstract

Due to the efforts of different authorities in the fight against illegal activities in the Tor networks, the traders have developed new ways of circumventing the monitoring tools used to obtain evidence of said activities. In particular, embedding textual content into graphical objects avoids that text analysis, using Natural Language Processing (NLP) algorithms, can be used for watching such onion web contents. In this paper, we present a Text Spotting framework dedicated to detecting and recognizing textual information within images hosted in onion domains. We found that the Connectionist Text Proposal Network and Convolutional Recurrent Neural Network achieve 0.57 F-Measure when running the combined pipeline on a subset of 100 images labeled manually obtained from TOIC dataset. We also identified the parameters that have a critical influence on the Text Spotting results. The proposed technique might support tools to help the authorities in detecting these activities.

Detection and formal classification of certainty and its application to text mining of chains of scholarly statements

“…7-The expression of Oct4 in various forms of human cancer [8], [9] and a recently described role for Oct4 in adult stem cells [10] and the expansion of epithelial progenitor cells [7] supports the theory that cancer is a disease of stem cells. This theory postulates that cancers arise in stem cells or early committed progenitors [58] due to their inability to differentiate in a regulated fashion.…”

“…With this in-mind, the observation that scientists' personal biases are increasingly being reflected in scholarly publications [7], [8] is troubling, as there is a breakdown in fidelity between the creation of a potentially biased scholarly assertion, the computational extraction of that assertion including its subtle bias, and the final consumption and interpretation of that extracted knowledge by another human. In response, a wide range of initiatives related to NLP have originated in recent years that attempt to improve the extraction of knowledge from scholarly text, such as detecting textual information in images [9], recognition of reference scope [10] or automatic assignment of perceived certainty [11]. These latter two are of particular relevance to this thesis.…”

“…Each sample starts in its own cluster and consecutively they are combined with other samples creating larger clusters. It is normally represented by a tree where leaves are the samples and the root is a single cluster gathering all samples inside 9 . In this thesis, to measure the distance between clusters, we employed the method 'average', where distance is calculated as the average of all the points in a cluster to all points in the other cluster [71].…”