Research on the impact of optimization algorithms on the accuracy of YOLOv11 neural networks

Abstract

Visual inspection and positioning based on image detection results is a rapidly growing component of automation systems. Machine vision is increasingly used in production lines for various technological purposes, as well as in special equipment. Improving recognition accuracy in such applications can be a difficult task, especially in conditions of possible limitations, one of which may be size and weight restrictions, which in turn limit the power of computer devices that implement image detection and recognition. A possible solution to this problem is to improve recognition accuracy by automatically tuning the hyperparameters of detection models using various optimization methods. This article presents the results of a study of the effectiveness of algorithms for automatically tuning the hyperparameters of the YOLO (You Only Look Once) image detection model, built on the basis of the SGD (stochastic gradient descent), Adam (adaptive learning rate estimation), and AdamW (improved version of Adam) optimization methods. The models were trained on the COCO 2017 dataset, limited to eight classes and balanced in terms of the number of images. For each optimization algorithm, 10 iterations of automatic selection were used, followed by training for 30 epochs. The test results showed that in tasks where detection accuracy is important and there are sufficient computational resources for hyperparameter selection during model training, it is advisable to use the stochastic gradient descent optimization algorithm, since the detection model configured with its use provides a higher probability of successful image recognition in real time under conditions of limited computational resources.