Análisis de algoritmos optimizadores de función de coste en el rendimiento de una red neuronal convolucional YOLO aplicada a la detección automática de peatones

Abstract

In recent years, research in the discipline of Deep Learning has improved great relevance in various real-life applications, including autonomous driving and object detection associated with this area. To carry out this process, it is necessary to train convolutional neural networks capable of performing fast and accurate detections. This work presents a comparative analysis of the performance of the YOLOv5 network, trained with three different cost function optimizers: Adam, SGD, and AdamW. The cost function optimization algorithms play a fundamental role in the training and performance of neural networks, as the cost function quantifies the discrepancy between actual values and values predicted by the model. The experiments were run using the Self Driving Car Dataset from Udacity and the X-Large version of YOLO. For the comparative analysis, the Kruskal-Wallis multiple comparisons test and the Dunn post hoc test were performed, from which the conclusion is drawn that the SGD optimizer differs significantly from Adam and AdamW. Although the results of a DL model are not generalizable, the results show that SGD achieved higher performance metrics than Adam and AdamW, achieving an average precision greater than 90%. These results highlight the importance of carefully selecting the cost function optimizer in the training of convolutional neural networks, such as YOLOv5, to improve performance and accuracy in object detection tasks, such as in the context of autonomous driving.