Evaluation of Deep Learning Strategies for Underwater Object Search

Underwater exploration based on underwater visual object detection suffers from many problems. The first one is definition of objects to be searched, as well as preparation of a sufficient number of their examples for classifier training. The other problem are harsh underwater conditions, such as water turbidity, noise, color vanishing, to name a few. In this paper we experiment with the deep learning approach, based on various convolutional neural networks, applied to the underwater object detection. Based on our underwater test sequences we conducted experiments aimed at measuring networks responses when trained on large databases of everyday objects when applied to underwater environment. In this paper new plausible-positive metric is introduced and experimental results, as well as conclusions and further directions are presented.