Splash Detection in Fish Plants Surveillance Videos Using Deep Learning

Abstract

The objective of this paper is to present and evaluate an improved method for automatic splash detection in surveillance videos of offshore fish production plants. In fishing and aquaculture industry one of the main challenges is production loss, that is, among the other things, caused by poor handling of the fish during operations such as crowding and delousing. This operations are very stressful for fish, and may trigger an increase in mortality, which is directly correlated with the production and profit loss. Because of this, improved solutions based on new technologies are being investigated, in order to decrease the risk of unnecessary stress, and improve the quality of production. One of the main parameters used for remote visual inspection of fish state is surface activity, which can be observed in a form of fish jumping and splashing. For that reason, in this paper, a novel algorithm based on using of Convolutional Neural Networks (CNNs) for splash detection is presented, which outperforms all existing algorithms based on local descriptors and linear classifiers. Using this approach we obtained splash detection accuracy of 99.9%.