Introduction
Nowadays, biometric systems have numerous applications in different fields. In recent years various applications of modern computer vision systems have been applied for different tasks like object detection, segmentation, and classification for analysing different fish vertebrates and aquatic animals. There are a lot of applications available for using artificial intelligence automated manners in different research fields of aquaculture studies, like fish welfare, disease detection, and fish behaviour. Fish species identification is the major tool for biologists’ studies to identify and trace these types of species over a period of time. Since machine learning can provide an automated manner to identify each individual fish, these methods can be much easier to use especially for real condition identification tasks. New machine learning approaches, such as different deep neural network architectures, could provide essential and more trustable tools for researching and developing the fisheries industry. Thus, in this article, we present a comparison of different convolutional neural neteorks (CNN) architectures in the field of individual image-based fish identification. The results of different techniques are discussed. Finally, we provided conclusions from applying different CNN methods of fish identification.
Materials and methods
In recent years numerous artificial neural network architectures have been introduced for different tasks like image classification and recognition. CNN have been used extensively in different image classification tasks like fish species identification. In this research, we used three different CNN architectures for performance review for fish individual identification tasks for Atlantic Salmon.Atlantic salmon identification – The dataset consists of 30 individual Atlantic salmon. The average fish weight was 251±21 g, and the length was 29.5±2.5 cm. The age of the fish was five months. All fish were tagged with PIT (passive integrated transponder) tags. The tagged fish were cultivated in a 2m3 recirculation freshwater tank for six months. Every two months, 4-8 images per fish were taken for each individual (session). Two to four images for each fish from each session were used to create the reference database for identification. All sessions were mixed together to identify the fish at different growth stages. The other images for each individual were used to evaluate the identification task.
We used simple CNN, VGG 16 and ResNet-50 for fish identification tasks. Stacked colour images of all individual fishes were used as the CNNs input for training and testing. A simple CNN can provide an accurate model over simple visual data; However, the pattern of salmon is more complex and change over time, so we need to use deeper networks to train a more accurate model for fish identification. Training different models with ResNet-50 with 23 million trainable parameters and VGG-16 with over 138 million trainable parameters show that using a deeper network can effectively handle the fish identification task for long-term identification. Below you can see the two deep networks used for fish identification.
Results
Fish identification using CNN: A simple CNN architecture with four hidden layers applied over the collected data, and the accuracy rate for identifying the 30 individual fish was 93.4% which corresponds to the high accuracy of individual identification.
Fish identification using VGGNet: The accuracy rate for identifying the 30 individual fish was 95.1%, which shows a better performance than simple CNN for fish identification.
Fish identification using ResNet-50: Finally, the accuracy rate for identifying the 30 individual fish using ResNet-50 using triplet loss function was 99.3%, which shows the best efficiency among the used CNN-based structures.
Conclusion
Identifying individual fish species and other aquatic species by tagging them is one of the most important tools for following up on their footprint in the aquatic habitat to estimate their health, growth assessment, and welfare for the fishing industry. This study showed the ability to use different CNN architectures on the visual data for individual fish identification. The three CNN architectures were evaluated on the salmon dataset. The best metrics result reported for the fish identification by applying ResNet-50 using triplet loss. A high accuracy rate of about 99% showed that it shows the effectiveness of applying this ResNet-50 architecture for fish identification tasks.
Acknowledgement: European Union’s Horizon 2020 research and innovation program under grant agreement No. 652831" (Aquauexcel2020), and GAJU 114/2022/Z
Reference:
Sébastien Villon, David Mouillot, Marc Chaumont, Emily S. Darling, Gérard Subsol, Thomas Claverie, Sébastien Villéger,A Deep learning method for accurate and fast identification of coral reef fishes in underwater images, Ecological Informatics, Volume 48, 2018, Pages 238-244, ISSN 1574-9541
Nuria Gómez-Vargas, Alexandre Alonso-Fernández, Rafael Blanquero, Luis T. Antelo, Re-identification of fish individuals of undulate skate via deep learning within a few-shot context, Ecological Informatics, Volume 75, 2023, 102036, ISSN 1574-9541
Lipeng Li, Feipeng Shi, Chunxu Wang, Fish image recognition method based on multi-layer feature fusion convolutional network, Ecological Informatics, Volume 72, 2022, 101873, ISSN 1574-9541