A unified pipeline for traffic sign detection using resolution aware dataset reduction

Abstract

Robust traffic sign detection is an important factor of safe autonomous driving, allowing vehicles to accurately interpret and comply with dynamic road regulations. However, real world deployment remains heavily challenged by distant, extra-small signs, as well as partial occlusions from trees, structures, or other vehicles. While optimizing training datasets through size reduction can dramatically lower computational overhead and storage footprints, conventional pruning methods often discard critical spatial features, severely degrading model accuracy. In light of these limitations, this paper introduces a novel, end to end processing pipeline that simultaneously tackles dataset size and small object detection bottlenecks. First, a lightweight VGG16 model is trained on representative samples to evaluate scene complexity patterns and dynamically predict an image specific compression threshold. Using this prediction, an adaptive, multi scale compression mechanism is applied exclusively to images that require it. A saliency guided image processing strategy is used to selectively compress background regions and preserve important Regions of Interest. This resolution aware dataset is then used to train the proposed Edge-Texture Fusion (ETF) YOLO11 network. By incorporating attention mechanisms to capture low-level edge and texture information, ETFYOLO11 improves the detection of distant and heavily compressed traffic signs. Three diverse publicly available traffic sign datasets are used to validate the effectiveness of the proposed method. This method reduces dataset storage requirements by 50–90%, cuts model parameters by 50%, and lowers GFLOPs by 20%.and at the same time achieve a higher mAP@50 than the baseline YOLO11n model.