Publication

Abstract : Over the last decade, extensive research has been conducted in managing and monitoring fishes in underwater environments. These researches aim to build solution that improves research on oceanography, marine biology and mainly for aquaculture fisheries. Aquaculture systems being controlled environments can boost production through regular monitoring of the conditions. Regular manual monitoring would be labor intensive and time efficient whereas traditional sensor-based solutions fail to be robust with respect to the highly complex underwater conditions. Traditional computer vision and image processing solutions fail due to challenges such as low luminosity, color imbalance and high image degradation. In this paper a robust deep learning based fish detection model is built using the state-of-the-art object detection algorithms YOLOv11 and YOLOv12 that are trained on a combination of high resolution and robust datasets DeepFish and OzFish in total containing 8275 images. Three different models from both the selected algorithms that varies by architectural complexity are trained and tested to analyzed. YOLOv11m achieved mean average precision of 78.3%, but YOLOv12s plotted the optimal balance between the model complexity and performance by achieving 77.6% with approximately half the model complexity.