We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Smoke semantic segmentation with multi-scale residual paths and weighted middle surveillances.
- Authors
Yuan, Feiniu; Zhang, Lin; Xia, Xue
- Abstract
Visual smoke segmentation is widely used for fire detection, simulation, human evacuation and pollution monitoring. However, it is challenging to accurately separate smoke from a single image due to highly complicated appearance of smoke, such as blurriness, semi-transparency and varying shapes. We design several multi-resolution convolutional paths to generate multi-scale feature maps for obtaining scale invariance. To capture the feature representation of blurriness and semi-transparency features, we present a Multi-scale Residual Module (MRM) by significantly increasing the width and depth of residual paths. Combining learnable de-convolution and bilinear interpolation has the advantage of generating multi-scale features that are helpful to progressively produce up-sampled and middle predictions for training surveillances. By designing a monotonically decreasing function with respect to the independent pre-training error of each prediction, we implement an automatic regulation of relative importance for accelerating the training. Experimental results show that our method achieves the best results of 79.03%, 78.30% and 78.65% on the three virtual smoke data sets among state-of-the-art methods. In addition, another superiority of our method is the proposed multi-prediction loss making the training of our network to converge faster and more stably.
- Subjects
SMOKE; CONVOLUTIONAL neural networks; POLLUTION monitoring
- Publication
Multimedia Tools & Applications, 2024, Vol 83, Issue 16, p47199
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-17260-2