Introduction
Optimizing fish production in intensive aquaculture can lead to increased production with respect to fish welfare. The new technologies enable to development of the Precision Fish Farming (PFF) (Fore et al. 2017) concept, whose aim is to apply control-engineering principles to fish production, thereby improving the farmer’s ability to monitor, control and document biological processes in fish farms. Individual fish identification is one of the tasks necessary for precision fish farming.
The widespread and popular methods of fish identification are tagging and marking (PIT, RFID or VIE tags), which are invasive methods. It was shown that image-based fish individual identification could substitute fish tagging (Bekkozhayeva 2021, Schraml 2020, Bekkozhayeva 2022 and Cisar 2021) . All the studies demonstrated long-term stability and high accuracy of image-based identification but only under controlled laboratory conditions. In this study, we implement image-based individual fish identification for real-time usage during the standard fish sampling procedure. The software used the ordinary web camera and was tested on rainbow trout identification with an accuracy of 100%.
Materials and methods
Rainbow trout Oncorhynchus mykiss (Fig.1 (right image)), a commercially important fish species, was used in our study. The size of the fish was around 13.7 ± 9.1 cm 32.8 ± 6.0 g. This species has a skin dot pattern on the body (juvenile pattern), which was used for the individual identification procedure. Those dots are different in size and location, which is expected to provide high discriminative power for individual identification within a close group of fish (tank). Fish were bought at the fish farm in Rybářství Litomyšl s.r.o. , Czech Republic. Finally, 1602 individuals were used in the experiment. Three pictures of the lateral view of the fish were collected for each individual at different angles and positions. An image of the right side of the fish was captured for all fish. The average resolution belonging to the one fish was 820 x 180 pixels.
Th e fish detection is based on the CNN with YOLOv8 architecture. YOLOv8 is one step bounding box detector . The CNN was trained using 187 manually labelled images of the fish randomly selected from 1602 fish. The original implementation from Ultralytics was used. The CNN was trained using Google colab environment. The trained CNN was then used in C# .net software developed by the authors for fish identification. The example of detected fish is in Fig. 1 .
The CNN performs the detection of the fish bounding box without the tail. The region of interest (ROI) used for identification is extracted from the central part of the fish. The area is defined by the percentage of fish wid th and height. The ROI is parametrized using a Histogram of oriented gradients descriptor to calculate the fish feature vector. The identification is implemented as nearest neighbou r classification, where the distance of the unknown fish is measured to all images in the database. The unknown fish is identified as the fish with the smallest distance. Two images of each fish were used as reference database and one image for each fish was used to test the identification accuracy.
Results
The accuracy of fish detection using the CNN approach was 100%. The accuracy of individual identification using nearest neighbour classification was 100%. The time of fish individual detection using CPU only (no GPU acceleration is needed) was 500ms. The time of fish identification of one fish within 1602 individuals was 800ms.
Conclusion
Individual fish identification is one of the keystones of the emerging concept of Precision Fish Farming. It was already proved by several studies that individual fish identification based on the fish skin pattern could substitute invasive fish tagging. In this paper, we developed the software for real-time fish individual identification using an ordinary web camera to provide support for fish sampling. The software was tested on 1602 individuals of juvenile rainbow trout and reached an accuracy of 100% . The software can be used by researchers or by fish farmers to speed up non-invasive fish sampling. The adaptation for other species is simple and needs just training CNN detectors for the new species.
Acknowledgement: European Union’s Horizon 2020 research and innovation program under grant agreement No. 652831" (Aquauexcel2020), GAJU 114/2022/Z
Reference:
Bekkozhayeva, D., Saberioon, M. & Cisar, P., 2021. Automatic individual non-invasive photo-identification of fish (Sumatra barb Puntigrus tetrazona) using visible patterns on a body. Aquacult Internationl .
Schraml, Rudolf & Hofbauer, Heinz & Jalilian , Ehsaneddin & Bekkozhayeva, Dinara & Saberioon, Mehdi & Cisar, Petr & Uhl, Andreas. 2020. Towards fish individuality-based aquaculture. IEEE Transactions on Industrial Informatics .
Bekkozhayeva, D., Císař, P., 2022. Image-Based Automatic Individual Identification of Fish without Obvious Patterns on the Body (Scale Pattern). Applied Sciences 12: 5401. (IF 2020 = 2,679; AIS 2020 = 0,409) https://doi.org/10.3390/app12115401
Císař, P., Bekkkozhayeva, D., Movchan, O., Saberioon, M., Schraml, R., 2021. Computer vision based individual fish identification using skin dot pattern. Scientific Reports 11: 16904. (IF 2020 = 4,379; AIS 2020 = 1,285) https://doi.org/10.1038/s41598-021-96476-4