Abstract:

This study proposes a two-level method for object detection aimed at annotating Arabic road sign content. Initially, the performance of three leading algorithms—YOLO, Faster R-CNN, and RetinaNet is evaluated. The research then focuses on overcoming challenges such as small symbol sizes and false panels by refining the YOLO algorithm through parameter optimization. To enhance detection capabilities further, the study incorporates deep active learning, enabling iterative improvements and dynamic training. The approach is tested on the Asayar dataset, which includes four sub-datasets covering panel, word-level, line-level, and symbol detection. The modified YOLOv8 algorithm demonstrates outstanding performance, achieving a mean Average Precision (mAP) of 0.94 for all objects and 0.98 for panel detection. These results underscore the method's ability to tackle the specific complexities of annotating Arabic and Latin text. This research contributes to the development of more effective object detection techniques and holds promise for advancing automated systems in traffic sign recognition and interpretation.