Deep Learning-Based Semantic Image Segmentation for Aquaculture

Strategic Insights -

Researchers at the Universidad Nacional del Altiplano in Peru have developed a novel approach to monitoring rainbow trout (Oncorhynchus mykiss) using deep learning-based semantic image segmentation. This breakthrough technology enables accurate size estimations and health monitoring, paving the way for sustainable aquaculture practices. The study employed two deep learning architectures, UNET and UNETR, to analyze high-resolution images of trout and compare their performance. The results show that UNET is more effective for aquaculture image segmentation, while UNETR requires more data to achieve satisfactory performance.

Key Takeaways:

  • The study utilized deep learning-based semantic image segmentation for enhanced monitoring of rainbow trout in Puno, Peru, achieving an IOU of 0.942854 with UNET and 256 x 256 pixel images after 20 epochs.
  • Experiment 2, using UNET with 512 x 512 pixel images, resulted in an IOU of 0.803244 after 50 epochs, indicating satisfactory performance despite increased complexity.
  • Experiment 3, employing UNETR with 256 x 256 pixel images, yielded lower IOU scores, with a best IOU of 0.253928, highlighting the challenge of training Transformer-based models with limited data.
  • The use of a coin as a scale reference in all experiments enabled precise conversion of pixel measurements to physical dimensions, allowing for accurate fish size estimations.
  • The study concluded that the results underscore UNET's effectiveness for aquaculture image segmentation, while also emphasizing data requirements for UNETR.
  • The approach provides a non-invasive, automated method for monitoring fish growth and health, contributing to sustainable aquaculture practices.

Statistics:

  • The study analyzed a dataset of 1200 high-resolution images.
  • UNET achieved an IOU of 0.942854 after 20 epochs, while UNETR yielded a best IOU of 0.253928.
  • Experiment 2 used L1Loss and Adam, resulting in an IOU of 0.803244 after 50 epochs.
  • The use of a coin as a scale reference enabled precise conversion of pixel measurements to physical dimensions.

Sources:

  • NewsRx. Universidad Nacional del Altiplano Researchers Report Research in Applied Sciences (Image Segmentation and Measurement of Trout Using a Convolutional Neural Network and Transformer Architecture). Science Letter. July 11, 2025; p 1791.
  • Image Segmentation and Measurement of Trout Using a Convolutional Neural Network and Transformer Architecture. Applied Sciences, 2025, 15(12):6873. (Applied Sciences - http://www.mdpi.com/journal/applsci)
  • Universidad Nacional del Altiplano
  • MDPI AG
  • Jose Cruz, E.P. Ingenieria Electronica, Universidad Nacional del Altiplano, Puno 21002, Peru.