Towards Automated Fish Detection Using Convolutional Neural Networks

Fisheries independent data one of the most important sources of information for fish stock assessments. Historically these data have been collected by a tools such as bottom trawls which are not effective or desirable in rocky or protected areas, In the last decade we have made significant progress in terms of using robotic platforms[1] [2] to collect optical imagery to assess fish stocks. We now routinely collect hundreds of thousands of images over a single research expedition. Fisheries biologists are overwhelmed by the large datasets that are being collected. In this paper we look at Convolutional Neural Networks [3] [4] as a mechanism to automatically detect and classify fish in underwater imagery. We present the results of analyzing a large dataset of underwater imagery comprising 10,000 images taken by the Seabed Autonomous Underwater Vehicle. The data is diverse - across different habitats, it exhibits no rotational symmetry, has large shadows compared to the organisms under consideration and also has large occlusions and objects that are small and not centered compared to the overall field of view. Despite these serious differences compared to land based image datasets we show that our segmentation and classification results are similar to state of the art efforts associated with land based applications.