Year: 2025
Type: dissertation
Abstract: Anadromous migratory fish species are monitored during their passage through fish migration control stations, on rivers. This trapping process allows the collection of biometric data and the description of external pathologies observed on the fish’s body, in order to study population’s health trends of these vulnerable species. These diseases and parasites are recorded using pathology codes, which compile the nature, location on the body, and severity of each damage. To meet the need for automating the detection of physical damage, the PathologIA project aims to apply deep leaning-based image analysis tools. The use of pre-trained models, along with supervised training models, has enabled the extraction and processing of information from photographs.The established pipeline is designed to detect pathologies, to segment the fish’s body into different morphological units, and characterize the location and severity of each identified pathology. Depending on whether the goal is to maximize comprehensiveness or caution in predicting the presence or absence of a pathology, a threshold must be selected for each pathology. This method enables the estimation of pathology codes with promising and useable quality for field operators. Any lack of precision could potentially be offset by supplementing the image database and labels used for the training process of the model.
Source: HAL
Author: Lecomte, Titouan
Institution: Institut Agro Rennes-Angers
Language: French
Subfield: Water Science and Technology
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7140957762