Super-resolved thermal imagery for high-accuracy facial areas detection and analysis - Publication - MOST Wiedzy

Search

Super-resolved thermal imagery for high-accuracy facial areas detection and analysis

Abstract

In this study, we evaluate various Convolutional Neural Networks based Super-Resolution (SR) modelsto improve facial areas detection in thermal images. In particular, we analyze the influence of selectedspatiotemporal properties of thermal image sequences on detection accuracy. For this purpose, a thermal facedatabase was acquired for 40 volunteers. Contrary to most of existing thermal databases of faces, we publishour dataset in a raw, original format (14-bit depth) to preserve all important details. In our experiments, weutilize two metrics usually used for image enhancement evaluation: Peak-Signal-to-Noise Ratio (PSNR) andStructural Similarity Index Metric (SSIM). In addition, we present how to design a SR network with a widenedreceptive field to mitigate the problem of contextual information being spread over larger image regions dueto the heat flow in thermal images. Finally, we determine whether there is a relation between achieved PSNRand accuracy of facial areas detection that can be analyzed for vital signs extraction (e.g. nostril region). Theperformed evaluation showed that PSNR can be improved even by 60% if full bit depth resolution data is usedinstead of 8 bits. Also, we showed that the application of image enhancement solution is necessary for lowresolution images to achieve a satisfactory accuracy of object detection.

Citations

  • 0

    CrossRef

  • 0

    Web of Science

  • 0

    Scopus

Details

Category:
Articles
Type:
artykuły w czasopismach
Published in:
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE no. 87,
ISSN: 0952-1976
Language:
English
Publication year:
2020
Bibliographic description:
Kwaśniewska A., Rumiński J., Szankin M., Kaczmarek M.: Super-resolved thermal imagery for high-accuracy facial areas detection and analysis// ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE -Vol. 87, (2020), s.103263-
DOI:
Digital Object Identifier (open in new tab) 10.1016/j.engappai.2019.103263
Verified by:
Gdańsk University of Technology

seen 8 times

Recommended for you

Meta Tags