A Study on Automatic Detection, Length and Mass Estimation of Fishes in Aqua Farming Environment

Biswas, R., Khonsari, R., Mutz, M. & Werth, D.(2024)
In: Santosh, K., et al. Recent Trends in Image Processing and Pattern Recognition. RTIP2R 2023. Communications in Computer and Information Science, vol 2026, S. 324-335, DOI: https://doi.org/10.1007/978-3-031-53082-1_26

Link zur Publikation:


Knowledge on body-length and mass of farmed fishes are essential for making informed business decisions in aqua-farming. These reduce wastage and promote sustainability that is critical for ecological preservation. In this paper, we present an approach for automatically determining the length and mass of fishes from images taken by a single camera placed above the fish-tank. Our motivation is to develop a solution that requires minimal hardware requirement and expert supervision for wide applicability. This is in sharp contrast with existing work in this area that employ complex hardware setup, e.g., stereo cameras, underwater camera, submerged tubes with twin-camera setup. These require technical expertise which ideally is not available to small businesses. Furthermore, our approach works with live fish in water compared to other techniques that work on dead fish. We evaluate different variants of the Mask-RCNN and YOLO algorithms for automatic fish detection. Subsequently, we use B-splines for estimating the length of fish in the pixel domain followed by physical length determination using optics. We obtain a mean average precision of 0.896 for the bounding box and a mean average precision of 0.902 for mask-based detection of fish with the YOLO model. The average error in physical length estimation is 3.32%. Lastly, we estimate the mass of fish by fitting the Length-Weight relationship using regression and obtain an average error of 12.24% in weight determination.

A Study on Automatic Detection, Length and Mass Estimation of Fishes in Aqua Farming Environment | SpringerLink