Publication

Abstract : The paper explores YOLO hyperparameter optimization crucial for robust deep learning models. It delves into optimizing parameters for three YOLO object detection categories: numeric, healthcare, and environmental. Dataset testing includes hand joint, plant, and numbers datasets. For healthcare, HSV augmentation improves generalization, yet altering color ranges might hinder diagnostic value detection. Geometric augmentation alterations can misinterpret image relationships. Adams optimizer excels for small datasets, yielding high precision and recall for all categories. A medium learning rate of 0.01 proves optimal, with deviations diminishing performance. Notably, a high learning rate impacts numeric detection negatively. Augmentation parameters like flip-up, flipdown, and translate can confuse digit detection; setting them to 0.0 improves accuracy, especially for numbers. The combination of a learning rate of 0.01, flip-up, flip-down, and translate at 0.0, with Adams optimizer, achieves superior precision and recall of 0.986 and 0.997 for numeric detection. The study underscores the significance of hyperparameter optimization in diverse object detection tasks, offering insights into parameter selection for optimal model performance.