FinFindR: Automated recognition and identification of marine mammal dorsal fins using residual convolutional neural networks. Issue 1 (19th July 2021)

• Record Type: Journal Article
• Authors: Thompson, Jaime W.; Zero, Victoria H.; Schwacke, Lori H.; Speakman, Todd R.; Quigley, Brian M.; Morey, Jeanine S.; McDonald, Trent L.
• Is Part Of: Marine mammal science. Volume 38:Issue 1(2022)
• Page Start: 139
• Page End: 150
• Publication Date: 2021-07-19
• Subjects: automated detection -- cetacean -- dolphin -- machine learning -- neural network -- noninvasive sampling -- photo‐identification -- Tursiops truncatus; Marine mammals -- Congresses; Marine mammals -- Periodicals; Marine mammals, Fossil -- Periodicals; Mammifères marins -- Périodiques; 599.5
• DOI: 10.1111/mms.12849
• Languages: English
• ISSNs: 0824-0469

Abstract

Abstract: Photographic identification is an essential research and management tool for marine mammal scientists. However, manual identification of individuals is time‐consuming. To shorten processing times, we developed finFindR, an open‐source application that uses a series of neural networks to autonomously locate dorsal fins in unedited field images, quantify an individual's unique fin characteristics, and match them to an existing photograph catalog. During a blind test comparing manual searching to finFindR for common bottlenose dolphin ( Tursiops Tursiops truncatus ) photographs, experienced photo‐identification technicians achieved similar match rates but examined an order of magnitude fewer photographs using finFindR (an average of 10 required with finFindR versus 124 with manual search). In those tests, the correct identity was ranked in the first position in 88% of cases and was within the top 50 ranked positions in 97% of cases. Our observations suggest that finFindR's matching capabilities are robust to moderate variation in image quality and fin distinctiveness. Importantly, finFindR allows users to build a catalog of known individuals through time and match an unlimited number of individuals instead of being restricted to a predefined set. finFindR 's convolutional neural networks could be re‐trained to identify members of many marine mammal species without altering finFindR 's inherent structure.