finFindR: Computer-assisted Recognition and Identification of Bottlenose Dolphin Photos in R
Thompson, J. W.; Zero, V. H.; Schwacke, L. H.; Speakman, T. R.; Quigley, B. M.; Morey, J. S. M.; McDonald, T. L.
Show abstract
O_LIPhotographic identification is an essential research and management tool for studying population size and dynamics of common bottlenose dolphins (Tursiops truncatus). Photographic identification involves recognizing individuals based on unique dorsal fin markings. Manual identification of dolphins, while successful, is labor-intensive and time-consuming. To shorten processing times, we developed a series of neural networks that finds fins, assesses their unique characteristics, and matches them to an existing catalog.\nC_LIO_LIOur software, finFindR, shortens photo-ID processing times by autonomously finding and isolating (i.e., \"cropping\") dolphin fins in raw field photographs, tracing the trailing edge of fins in cropped images, and producing a sorted list of likely identities from a catalog of known individuals. The program then presents users with the top 50 most likely matching identities, allowing users to view side-by-side image pairs and make final identity determinations.\nC_LIO_LIDuring testing on two sets of novel images, finFindR placed the correct individual in the first position of its ordered list in 88% (238/272 and 354/400) of test cases. finFindR placed the correct identity among the top 10 ranked images in 94% of test cases, and among the top 50 ranked images in 97% of test cases. Hence, if a match does not exist in the first 50 images of finFindRs ordered list, researchers can be almost certain (~97%) that a match does not exist in the entire catalog.\nC_LIO_LIDuring a head-to-head blind test of the human-only and finFindR-assisted matching methods, two experienced photo-ID technicians both achieved 97% correct identification of identities when matched against a catalog containing over 2,000 known individuals. However, the manual-only technician examined 124 images on average before making a match, while the technician using finFindR examined only 10 images on average before finding a match.\nC_LIO_LIWe conclude that finFindR will facilitate equal or improved match accuracy while greatly reducing the number of examined photos. The faster matches, automated detection, and automated cropping afforded by finFindR will greatly reduce typical photo-ID processing times.\nC_LI
Matching journals
The top 6 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- ML-morph: A fast, accurate and general approach for automated detection and landmarking of biological structures in images 95%
- Deep learning-based methods for individual recognition in small birds 95%
- Thinking like a naturalist: enhancing computer vision of citizen science images by harnessing contextual data 95%
Similar papers in this journal
- Leveraging synthetic data produced from museum specimens to train adaptable species classification models 94%
- Accuracy and precision of citizen scientist animal counts from drone imagery 93%
- Automated data extraction from historical city directories the rise and fall of mid-century gas stations in Providence, RI 93%
Similar papers in this journal
Similar papers in this journal
- Understanding Ecological Systems Using Knowledge Graphs: An Application to Highly Pathogenic Avian Influenza 90%
- Batch-effect correction in single-cell RNA sequencing data using JIVE 89%
- The network makeup artist (NORMA-2.0): Distinguishing annotated groups in a network using innovative layout strategies 89%
Similar papers in this journal
- A benchmark dataset for individual tree crown delineation in co-registered airborne RGB, LiDAR and hyperspectral imagery from the National Ecological Observation Network 92%
- Bird song comparison using deep learning trained from avian perceptual judgments 92%
- Distinguishing shadows from surface boundaries using local achromatic cues 91%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.