RESUMO
Retrieving text embedded within images is a challenging task in real-world settings. Multiple problems such as low-resolution and the orientation of the text can hinder the extraction of information. These problems are common in environments such as Tor Darknet and Child Sexual Abuse images, where text extraction is crucial in the prevention of illegal activities. In this work, we evaluate eight text recognizers and, to increase the performance of text transcription, we combine these recognizers with rectification networks and super-resolution algorithms. We test our approach on four state-of-the-art and two custom datasets (TOICO-1K and Child Sexual Abuse (CSA)-text, based on text retrieved from Tor Darknet and Child Sexual Exploitation Material, respectively). We obtained a 0.3170 score of correctly recognized words in the TOICO-1K dataset when we combined Deep Convolutional Neural Networks (CNN) and rectification-based recognizers. For the CSA-text dataset, applying resolution enhancements achieved a final score of 0.6960. The highest performance increase was achieved on the ICDAR 2015 dataset, with an improvement of 4.83% when combining the MORAN recognizer and the Residual Dense resolution approach. We conclude that rectification outperforms super-resolution when applied separately, while their combination achieves the best average improvements in the chosen datasets.
RESUMO
An increasing number of musicians are opting to use tablet devices instead of traditional print media for their music sheets since the digital medium offers the benefit of storing a lot of music in a compact space. The limited screen size of the tablet devices makes the music difficult to read and musicians often opt to display part of the music page at a time. With fewer music lines on display, the musician will then have to resort to scrolling through the music to read the entire score. This scrolling is annoying since the musicians will need to remove their hands from the instrument to interact with the tablet, causing a break in the music if this is not done quickly enough, or if the tablet is not sufficiently responsive. In this paper, we describe an alternative page turning system which automates the page turning event of the musician. By actively monitoring the musician's on-screen point of regard, the system retains the musician in the loop and thus, the page turns are attuned to the musician's position on the score. By analysing the way the musician's gaze changes between attention to the score and the instrument as well as the way musicians fixate on different parts of the score, we note that musicians often look away from the score and toward their hands, or elsewhere, when playing the instrument. As a result, the eye regions fall outside the field-of-view of the eye-gaze tracker, giving rise to erratic page-turns. To counteract this problem, we create a gaze prediction model that uses Kalman filtering to predict where the musician would be looking on the score. We evaluate our hands-free page turning system using 15 different piano songs containing different levels of difficulty, various repeats, and which also required playing in different registers on the piano, thus, evaluating the applicability of the page-turner under different conditions. Performance of the page-turner was quantified through the number of correct page turns, the number of delayed page turns, and the number of mistaken page turns. Of the 289 page turns involved in the experiment, 98.3% were successfully executed, 1.7% were delayed, while no mistaken page turns were observed.