| Title |
Deep Contour Recovery: Repairing Breaks in Detected Contours using Deep Learning |
| Authors |
(Minyoung Kyoung) ; (Hyunbean Yi) |
| DOI |
https://doi.org/10.5573/IEIESPC.2020.9.10.371 |
| Keywords |
Contour recovery; Contour detection; Deep learning; CNN; Residual connections training |
| Abstract |
We present a contour recovery framework based on a deep learning model to connect broken contours (breaks) produced by contour detection methods. The idea is that the convolutional neural network iteratively predicts vectors that can grow along the direction of the true contour from the end points of the breaks. For this prediction, we use residual connections training, which models continuous predictions from the previous inference. However, conventional residual connections training is prone to gradually accumulating errors at each inference step. In this work, we propose a ground truth selection algorithm and sub-iteration training to efficiently and reliably train a deep learning model. The ground truth selection extracts a small set of coordinates to represent an actual contour. The sub-iteration training creates the next input that is predicted by additional training of a network replicated from the main network. Our experimental results demonstrate that the ground truth selection creates a ground truth suitable for contour recovery. Moreover, our approach improves the performance of contour detection when applied to the results of existing representative contour detection methods. |