Rendimiento de yolov8n en la comparación de algoritmos de deep learning para el conteo automático de postlarvas de colossoma macropomum

• Autor: Pereira Ramos, Piero
• Formato: tesis de grado
• Fecha de Publicación: 2024
• Institución: Universidad Nacional De La Amazonía Peruana
• Repositorio: UNAPIquitos-Institucional
• Lenguaje: español
• OAI Identifier: oai:repositorio.unapiquitos.edu.pe:20.500.12737/11287
• Enlace del recurso: https://hdl.handle.net/20.500.12737/11287
• Nivel de acceso: acceso abierto
• Materia: Algoritmos de detección; Visión por ordenador; Redes neuronales convolucionales; Detección de peces; Larvas de peces; Gamitana; Colossoma macropomum; https://purl.org/pe-repo/ocde/ford#2.02.04

This study aimed to compare the effectiveness of the YOLOv8n algorithm with different deep learning models in the task of automatic detection and counting of Colossoma macropomum postlarvae, a key process to improve aquaculture production and reduce mortality in this field. In order to evaluate the quality of the results, metrics such as accuracy, sensitivity, F1-Score and processing time were analyzed, comparing the performance of YOLOv8n with PP-PicoDet-det-L, Faster R-CNN and Grid R-CNN. The methodology employed included preprocessing and data augmentation techniques applied to a set of 71 images obtained from various mobile devices, which ensured greater representativeness and quality of the sample. The training of the algorithms was carried out in 12 epochs, using both a supercomputer and a workstation provided by IIAP. The results indicate that YOLOv8n exhibits superior performance in terms of detection and automatic counting, as it achieved higher values of precision, sensitivity and F1-Score, while registering a shorter processing time compared to the other algorithms. It should be noted that it achieved a counting time of only 56 seconds, surpassing the other models in terms of speed and efficiency. If future work is able to implement the YOLOv8n algorithm in mobile applications or embedded software, this research could provide a practical solution for the aquaculture industry by using deep learning to automate postlarvae counting, with the potential to improve operational efficiency and reduce costs.