Deep learning methods for processing endoscopic high-speed video and laryngeal parameter estimation

Více o knize

Deep learning methods have significantly impacted computer vision, image processing, and related fields. This dissertation investigates their application to enhance and process endoscopic high-speed video (HSV), a key technique in voice research due to the need for sophisticated recording of the rapid oscillation of vocal folds. Voice disorders negatively affect the quality of life for individuals and society, highlighting the need for more objective diagnostic techniques. The dissertation presents several contributions: an innovative method to enhance low-light HSV using an improved U-Net convolutional neural network; a robust, fast deep-learning-based automatic method for glottis segmentation in HSV data; development of an enhanced two-mass model of the vocal folds; and proof of concept for estimating ex-vivo subglottal pressure validated with experimental data, as well as estimating subglottal pressure using a recurrent neural network trained on a numerical model. Following a comprehensive introduction to voice research and deep learning, the dissertation details the developed methods and results, showcasing significant advancements in low-light image enhancement, automatic glottis segmentation, and physical voice parameter inference.