Unsupervised Feature Fusion Model for Marine Raft Aquaculture Sematic Segmentation Based on SAR Images

Mengmeng Li, Xinzhe Wang, Jianchao Fan

Abstract:

Marine aquaculture sematic segmentation provides
a scientific basis for marine regulation and plays an important
role in marine ecological protection and management. Currently,
most high-performance marine aquaculture segmentation networks are trained by supervised learning. This approach requires
collecting a large number of accurate manually labelled samples
for training, but the labelled samples are difficult to obtain. To
solve this problem, this paper proposes an unsupervised feature
fusion model (UFFM) for marine raft aquaculture semantic
segmentation. Firstly, a pseudo-label generator is designed to
label the training samples, and a coarse mask is generated using
saliency feature clustering. The training samples with pseudolabels are inputted into a multilevel feature fusion module to
further extract and continuously improve the graphical shapes
and categories of the objects under the guidance of cross-entropy
loss. The pseudo-labels are further optimised under continuous iteration to improve the model segmentation performance.
Comparison experiments on the GF-3 dataset demonstrate the
effectiveness of UFFM.

Submission Number: 72