05 / 2022 – 10 / 2023
Bats form the second largest group of mammals and are an important component of many ecosystems. Yet, we still know little about the life history of many bat species. This is mainly due to their nocturnal and fast lifestyle, which makes them difficult to observe and track . In addition, the animals are difficult to mark individually. Researchers usually opt for one of two options, both of which are quite invasive: transponders – injected under the skin – or rings on the forearm that often cause inflammations . Within this thesis we aim to develop a new, non-invasive identification method utilizing bat wing visuals and computer vision techniques. The recognition of a unique wing print requires the presence of a visible and invariable biometric feature. Potential candidates include collagen/elastin fibres inside bats’ wing that serve as stabilization against aerodynamic loads during flight  and form large-scale fingerprint-like patterns, blood vessels traversing the wing membrane, or scars of previous injuries. Previous work has already shown promising results when humans manually matched images of individuals using collagen fiber structures inside the wings for differentiation . As bat wings are particularly thin, both their external and internal structure can be captured by a camera filming against an infrared(IR)-illuminated background, making it a promising approach for automated re-identification of individuals.
Research coverage of automated re-identification among animal species is generally scarce. A lot of related work targets humans in a security context – e.g. facial recognition for law enforcement . However, the concept of biometric features is in general not species-dependent, therefore the same algorithms can be adapted to different input domains . Our goal is to build on this work and investigate the feasibility of using transfer learning based on a pretrained instance of a proven neural network architecture such as ResNet-50 to automate the process of re-identifying individual bats from photos of their wing membranes.
We want to develop the Wing Print method in collaboration with Nuremberg Zoo, which houses a big colony (>200 individuals) of Glossophaga mutica, a species of nectar feeding bats. Due to flowers being their main food source, they are able to hover in the air , so we can lure them into the designated position in a camera trap with nectar-like nutrient solution and then have a certain time frame available for taking photos with a high-speed camera. After acquiring suitable video hardware, we will engineer a setup that allows us to take pictures of the bat wings in flight with consistent quality under reproducible conditions. The hardware is going to be built around a feeder that acts as an artificial flower for the bats. An IR light barrier inside triggers a signal as soon as an animal approaches and inserts its snout into the opening, providing a nectar reward. The camera mount will have a gantry-like design around the feeder that combines an IR-illuminated background below the bat with a corresponding high speed camera on top. Once the setup is ready for a field test, regular recording sessions will be conducted at the zoo in order to generate raw film material. During preprocessing frames that do not show the bats’ wings completely unfolded must be sorted out – e.g. with the help of image masks and grey level histograms. Once the relevant images have been extracted and manually labelled according to their affiliation to individuals, we will take a pre-trained instance of ResNet-50 that has been trained on 1000 different classes of the IMAGENET image database as our baseline for transfer learning and replace its classifier with one suitable for closed-set classification of the individual animals in our bat data. Only the new classifier is then trained on these images. Building on a pre-trained network should save us a lot of training time and still give good results. After training, we will evaluate the network’s overall performance in re-identifying the bats by measuring the accuracy the network achieves when trying to re-identify the same individuals across several image sets captured at different times under different conditions. Potentially, we are also able to test against data from other sources, depending on its availability. This could be more pictures/videos of bats or data from different species incorporating similar biometric markers. Based on our findings we can then explore possible optimizations in classifier architecture, hyperparameters and image preprocessing to boost the network’s performance. Overall, the goal of this thesis is to develop a camera-based image capturing setup for bat wings and to use the gathered data to train a state-of-the-art Re-ID model so it can distinguish between individuals. A strongly associated goal is to examine and validate the robustness of the re-identification against labelled reference data.
 T. H. Kunz and S. Parsons, “Ecological and behavioral methods for the study of bats,” KIP Articles, no. 5833, 2009. [Online]. Available: https://digitalcommons.usf.edu/kip articles/5833
 B. Mellado, L. d. O. Carneiro, M. R. Nogueira, and L. R. Monteiro, “The impacts of marking on bats: mark-recapture models for assessing injury rates and tag loss,” Journal of Mammalogy, vol. 103, no. 1, pp. 100–110, 12 2021. [Online]. Available: https://doi.org/10.1093/jmammal/gyab153
 A. Hedenstr¨om and L. C. Johansson, “Bat flight: aerodynamics, kinematics and flight morphology,” Journal of Experimental Biology, vol. 218, no. 5, pp. 653–663, 03 2015. [Online]. Available: https://doi.org/10.1242/jeb.031203
 S. K. Amelon, S. E. Hooper, and K. M. Womack, “Bat wing biometrics: using collagen–elastin bundles in bat wings as a unique individual identifier,” Journal of Mammalogy, vol. 98, no. 3, pp. 744–751, 2017.
 S. Kumar, S. Singh, R. Singh, and A. Singh, Biometric Methods for Animal: Recent Trends and Future Challenges. Springer Singapore, 01 2017, pp. 223–243.
 X. Lin, P. Ren, C.-H. Yeh, L. Yao, A. Song, and X. Chang, “Gait recognition for person re-identification,” The Journal of Supercomputing, 2021.
 X. Shi, C. Yang, X. Xia, and X. Chai, “Deep cross-species feature learning for animal face recognition via residual interspecies equivariant network,” in Computer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 667–682.  J. S. Findley, E. H. Studier, and D. E. Wilson, “Morphologic properties of bat wings,” Journal of Mammalogy, vol. 53, no. 3, pp. 429–444, 1972. [Online]. Available: http://www.jstor.org/stable/1379035