Meeting Abstract
Field-based research projects utilizing automated image capture devices often rely on humans to identify images. This method of analysis is limited in that data can only be retroactively parsed by an observer looking to identify patterns in behavior. If birds could be identified in real-time by an automated process integrated into the camera device, not only would there be hours of time saved identifying animals, it would also enable the possibility of a varied response to different subjects at the time of data collection. We accomplished exactly this as part of the PASSER program by use of Tensor Flow, a neural networking package for the programming language Python that allows users to create a neural network for image classification. A neural network requires a large repository of photos when being used for image classification so that it can provide more accurate identification of the images it will be given. The PASSER project has collected hundreds of thousands of photos of various bird species to be fed into the neural network so that it can fine tune its “neurons” for identification of various species and sexes. The smart feeder can then be updated to use the neural network on its own so that the system sustains itself without human interaction other than necessary hardware maintenance. The neural network, with enough collected data, could potentially even be used to identify individual birds if that specific individual bird frequents the feeder sufficiently often. The uses for this data are abundant as patterns could be easily localized and analyzed without the necessity of countless man hours spent identifying birds. Processes like these may enable a shift of focus from simple species identification, into broad-based environmental-ethological analysis.
