YOLO-CBAM: ЛЕГКОВЕСНАЯ НЕЙРОСЕТЕВАЯ МОДЕЛЬ С ИНТЕГРИРОВАННЫМ МЕХАНИЗМОМ ВНИМАНИЯ ДЛЯ АНАЛИЗА ДОРОЖНЫХ ПОВРЕЖДЕНИЙ

Abstract

In this paper, we propose YOLO-CBAM, a fast and lightweight neural network for real-time road surface defect detection. We improved the standard YOLO architecture by adding a convolutional block attention module (CBAM). This attention mechanism helps the model focus on complex road defects, such as cracks, and ignore background visual noise. Experiments were conducted to test how different weight learning methods affect the model training process. The results show that using partial transfer learning helps the network quickly overcome early errors and achieve an average accuracy (mAP@50) of 0.6 in just 40 epochs. Consequently, the YOLO-CBAM model improves defect detection without incurring high computational costs, making it a highly effective solution for video surveillance systems in smart cities.