Authors:
Gordon Böer
1
;
Rajesh Veeramalli
1
and
Hauke Schramm
2
;
1
Affiliations:
1
Institute of Applied Computer Science, Kiel University of Applied Sciences, Kiel, Germany
;
2
Department of Computer Science, Faculty of Engineering, Kiel University, Germany
Keyword(s):
Semantic Segmentation, Underwater Imagery, Fish Detection.
Abstract:
The semantic segmentation of fish in real underwater scenes is a challenging task and an important prerequisite for various processing steps. With a good segmentation result, it becomes possible to automatically extract the fish contour and derive morphological features, both of which can be used for species identification and fish biomass assessment. In this work, two deep learning models, DeepLabV3 and PSPNet, are investigated for their applicability to fish segmentation for a fish stock monitoring application with low light cameras. By pruning these networks and employing a different encoder, they become more suitable for systems with limited hardware, such as remotely operated or autonomously operated underwater vehicles. Both segmentation models are trained and evaluated on a novel dataset of underwater images showing Gadus morhua in its natural behavior. On a challenging test set, which includes fish recorded at difficult visibility conditions, the PSPNet performs best, and ach
ieves an average pixel accuracy of 96.8% and an intersection-over-union between the predicted and the target mask of 73.8%. It achieves this with a very limited parameter set of 94,393 trainable parameters.
(More)