AQUA 2024

August 26 - 30, 2024

Copenhagen, Denmark

ADAPTING CNN-BASED INDIVIDUAL IDENTIFICATION TO IDENTIFY FISH INDIVIDUALS UNDER REAL CONDITIONS

Mohammad Mehdi Ziaei* , Petr Cisar

 

 Faculty of Fisheries and Protection of Waters, CENAKVA ,  USB, Ceské Budejovice , Zámek 136, Nové Hrady 373 33, Czech Republic

Email* : ziaei@frov.jcu.cz

 



The interest in identifying individual fish within populations has grown in the fields of aquaculture and fisheries management in recent years.  The common approach for individual identification is physical tagging, which can affect the fish’s growth and welfare .  Due to  the issues associated with invasive individual fish identification, non-invasive methods are progressively superseding traditional tagging techniques to minimize or eliminate direct interaction with the fish. In this research field, several computer vision-based methods have been developed for individual fish identification. However, these methods have not presented a generalized model for individual fish identification , particularly  for  long-term  applications and real-world conditions. The primary objective of the presented research was to investigate the capability of deep learning techniques as a generalized model for performing real-time individual fish identification task under real-world conditions.

For this study, a closed group of 34 individuals of rainbow trout  juveniles  (Oncorhynchus mykiss) were cultivated and kept in the fish tanks for over 6 months. After two months, all of the fish were PIT-tagged .  One minute video record of the fish with known ID swimming in the aquarium was recorded. The other fish were visible behind (the background scene) the transparent separator to simulate the real-condition environment. The video sampling was conducted using an RGB camera, capturing footage of  the aquarium at a resolution of 1280x720 pixels. The camera recorded at a frame rate of 80 frames per second, resulting in smooth motion.  The PIT tag reader read the tag of the front fish and recorded it as the corresponding ID. For each fish, 10 frames of video were extracted when the fish was oriented to the left side and had less movement. T he  upper  dorsal fin, eye  and fish body detector were trained from the dataset. The detectors were applied to extracted frames, and  the rectangular area of fish skin with the highest pigmentation pattern stability was automatically cropped as the region of interest for identification.  The dataset for individual identification task  consists of 680 images of regions of interest. The triplet network with resnet-50 arch itecture was used as  a  classifier to model  the  distinguishing features of individuals of fish.

 The performance of the trained identifier model was 92 % accuracy between  the last two sessions ( after 30 days of fish growth).  The evaluation results and performance of the  CNN-based  model showed that the proposed method is able to learn the long-term invariant  pigmentation patterns that provide  precise  identification of individual fish over a period of time in similar conditions for real-world application.