NorFisk: fish image dataset from Norwegian fish farms for species recognition using deep neural networks

Abstract

Long-term autonomous monitoring of wild fish populations surrounding fish farms can contribute to a better understanding of interactions between wild and farmed fish, which can have wide-ranging implications for disease transmission, stress in farmed fish, wild fish behavior and nutritional status, etc. The ability to monitor the presence of wild fish and its variability with time and space will improve our understanding of the dynamics of such interactions and the implications that follow. Automatic fish detection from video streams at farm sites using neural networks may be a suitable tool. However there are not many image datasets publicly available to train these neural networks, and even fewer that include species that are relevant for the aquaculture sector. This paper introduces the first version of our dataset, NorFisk, which can be found publicly available at Crescitelli (2020). It contains 3027 annotated images of saithe and 9487 of salmonids and it is expected to grow in the near future to include more species. Annotated image datasets are typically built manually and it is a highly time-consuming task. This paper also presents an approach to automate part of the process when generating these types of datasets with fish underwater. It combines techniques of image processing with deep neural networks to extract, label, and annotate images from video sources. The latter was used to produce NorFisk dataset by processing video footage taken in several fish farms in Norway.

NorFisk: fish image dataset from Norwegian fish farms for species recognition using deep neural networks