Research Information System
CLUSTER COMPUTING-THE JOURNAL OF NETWORKS SOFTWARE TOOLS AND APPLICATIONS, vol.27, no.9, pp.12787-12797, 2024 (SCI-Expanded)
Thermographic inspection is particularly effective in identifying thermal bridges because it visualizes temperature differences on the building's surface. The focus of this work is on energy-efficient computing for deep learning-based thermal bridge (anomaly) detection models. In this study, we concentrate on object detection-based models such as Mask R-CNN_FPN_50, Swin-T Transformer, and FSAF. We do benchmark tests on TBRR dataset with varying input sizes. To overcome the energy-efficient design, we apply optimizations such as compression, latency reduction, and pruning to these models. After our proposed improvements, the inference of the anomaly detection model, Mask R-CNN_FPN_50 with compression technique, is approximately 7.5% faster than the original. Also, more acceleration is observed in all models with increasing input size. Another criterion we focus on is total energy gain for optimized models. Swin-T transformer has the most inference energy gains for all input sizes ( approximate to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx$$\end{document} 27 J for 3000 x 4000 and approximate to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\approx$$\end{document} 14 J for 2400 x 3400). In conclusion, our study presents an optimization of size, weight, and power for vision-based anomaly detection for buildings.