| Title |
A fast algorithm for facilitating heartbeat annotation in long-term ECG signals / |
| Authors |
Santos Rodrigues, Ana ; Lukoševičius, Mantas ; Marozas, Vaidotas |
| DOI |
10.23919/CinC53138.2021.9662850 |
| ISBN |
9781665467216 |
| eISBN |
9781665479165 |
| Full Text |
|
| Is Part of |
Computing in cardiology (CinC): September 12-15, 2021, Brno, Czech Republic.. Piscataway, NJ : IEEE, 2021. p. 1-4.. ISSN 2325-8861. eISSN 2325-887X. ISBN 9781665467216. eISBN 9781665479165 |
| Keywords [eng] |
CNN ; deep learning ; pavement defects ; U-Net ; loss function |
| Abstract [eng] |
Convolutional neural networks perform impressively in computer vision for a complicated image segmentation tasks. In many cases, architectural adjustment decisions are engaged to improve model performance. In this research, we address pixel-level performance improvement by utilizing different single and combined model training loss functions. The convolution encoder-decoder U-Net is employed for pavement defects segmentation in CrackForest dataset. We also propose a modified weighted cross-entropy that reduces neural network penalization when an error occurs around defect edges. Statistical and visual performance evaluation is taken into consideration for model comparison. Up to 0.6914 Dice score is received and 0.0179 improvement can be achieved by only changing loss function during the same neural network architecture training process. The trained model process a 480×320 greyscale image in 32ms. |
| Published |
Piscataway, NJ : IEEE, 2021 |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2021 |
| CC license |
|
