Real-time Weakly Supervised Semantic Segmentation of Seabed Sediments in Side-scan Sonar Images
Text Complet
Compartir
Real-time Weakly Supervised Semantic Segmentation of Seabed Sediments in Side-scan Sonar Images
Distinguishing between marine benthic habitat characteristics is of key importance in a wide
array of applications from installations of oil rigs to laying networks of cables and
monitoring the impact of humans on marine ecosystems. The Side Scan Sonar (SSS) is a
widely used sensor in this regard. It works on the principle of acoustic propagation and
reflection to produce high-resolution images by logging the intensities of sound waves
reflected back from the seafloor.
The goal of this work would be to leverage these acoustic intensity maps to produce
pixel-wise categorization of different seafloor types.
The annotations of the seafloor used to supervise model training were somewhat noisy.
This results from the fact that the annotations were made on SSS mosaics while we are
working with raw SSS waterfalls. Transferring these annotations to raw waterfalls is not a
straightforward process, especially without having access to the internal parameters used
for mosaicing and thus, leads to certain discrepancies. Therefore, the ground truth
generated for the raw waterfalls is not pixel-wise accurate and the trained models suffer
from weak supervision. Therefore, we plan to adopt a weakly supervised learning
framework to achieve our goal of seabed segmentation. We further plan to supplement the
framework by leveraging the noisy ground truth that we have available acting as pseudo
masks to regularize training.
Steps to be done:
● The structure of available data should be understood
● Selection of two best approach to implement
● Two best approach should be working, tuned and trained (loss converging) on
remote server (Falcon)
● Comparison of performance of the two approach (speed, IoU)
● Comparison with fully supervised approach (chosen baseline)
keywords: Seafloor Segmentation, side-scan sonar, Weakly supervised approach, real-time.