Towards Robust Identification of Slow Moving Animals in Deep-Sea Imagery by Integrating Shape and Appearance Cues

This paper describes an approach for the detection of stationary crabs of various sizes in underwater images. Specific issues related to underwater imaging such as low contrast and non-uniform lighting are addressed by the pre-processing step. The segmentation step is based on colour, size, and shape considerations. Segmentation identifies regions that potentially correspond to crabs. The shape of these regions is de- scribed via moment-based feature vectors invariant to translation, size, and rotation, which are fed into a feed-forward neural network for classiffication purposes. Experimental results are promising and show that future work needs to focus on improving both the segmentation step and the training procedure for the neural network.