Desarrollo de una aplicación para el monitoreo del crecimiento de tilapias utilizando machine learning

Abstract

The application of machine learning has great potential in aquaculture as it can provide important data on fish growth that does not require invasive measures. Furthermore, the introduction of advanced techniques in aquaculture contributes to the technological advancement of the sector, which has a positive impact on economic indicators, sustainability and social well-being. The objective of the project lies in the need to apply agile and non-invasive estimates on the growth of tilapia that allow optimizing manual measurements, in addition to using statistical techniques to estimate the length and weight of tilapia, leaving aside traditional methods. To achieve this project, a series of steps were carried out. First, a literature review focused on weight estimation using computer vision was carried out. Second, about 1000 tilapia images were collected for the construction of the dataset. Third, the YOLOv8 pre-trained model was adapted for the segmentation of tilapia, in addition to creating a model using a statistical technique such as linear regression specifically for the estimation of weight and length. In this way, with both models, they were integrated into a web page using the Extreme Programming or XP methodology, in order to develop the web application with its phases: planning, design, coding and testing. Using a confusion matrix, the detection and segmentation model was evaluated with the precision and accuracy metrics, reaching a precision of 88%, in addition to evaluating the model with the ISO 9126 standard with the objective of assessing functionality, reliability, usability and efficiency. of this web application. At the end of the tests, it was successfully verified that this web application can detect, segment and predict both the length and weight of a tilapia using an image with a certain degree of precision to the real values, however, the predictions in some cases show a significant discrepancy, possibly due to limited variability in data collected for length and weight