Comparison of architectures and training strategies for convolutional neural networks intended for location-specific counting of slaughter pigs

Research output: Contribution to conferencePaperResearchpeer-review

Pen fouling is an undesired behaviour seen in growing pigs, where they start resting in the excretion area and excrete in the designated resting area. It is reasonable to assume that automatic monitoring of the location of the pigs within the pen could be used for early warnings of imminent pen fouling events. We intend to provide such automatic monitoring using convolutional neural networks (CNN) applied to images captured above the pens. In this preliminary study, we compare 12 different combinations of CNN architectures and training strategies for this purpose. The best performing strategy yielded an overall mean absolute error of 0.35 pigs and a coefficient of determination of 96 % between the predicted and observed number of pigs in a given area of the pen.
Original languageEnglish
Publication yearAug 2019
Number of pages8
Publication statusPublished - Aug 2019
EventThe 9th European Conference on Precision Livestock Farming - Cork, Ireland
Duration: 26 Aug 201929 Aug 2019

Conference

ConferenceThe 9th European Conference on Precision Livestock Farming
CountryIreland
CityCork
Period26/08/201929/08/2019

    Research areas

  • conventional neural network, fouling, monitoring, slaughter pig

See relations at Aarhus University Citationformats

ID: 163162082