FAIERDet: Fuzzy-based adaptive image enhancement for real-time traffic sign detection and recognition under varying light conditions

Abstract

Intelligent vehicle safety relies on the Traffic Sign Detection and Recognition (TSDR) system. The real-world varying light affects the visibility of traffic signs and emphasizes the necessity of a robust TSDR. Existing solutions face challenges in effectively balancing accuracy and inference time for such challenges. The proposed work introduces an adaptive framework integrating preprocessing and enhancement modules with the existing detection network. The preprocessing module employs a Fuzzy Inference System (FIS) to evaluate the illumination channel and calculate the image's exposure quality. Low-light images are directed to enhancement depending on the exposure quality, while good-light images are passed to the detection network directly. The enhancement module improves image brightness while preserving color details through illumination adjustment using the proposed Adjustment Factor Prediction Convolutional Neural Network (AFPCNN). Finally, YOLOv8 is used for TSDR from the image. The results entail accuracy comparisons of simulated low-light images using three publicly available datasets: the German Traffic Sign Detection Benchmark (GTSDB), the Tsinghua-Tencent 100K (TT100K), and the Mapillary Traffic Sign Dataset (MTSD). The proposed enhancement module improves Recall and mean Average Precision on randomly dark images by 10–18 % and 5–9 % across the benchmark datasets. Furthermore, the proposed framework enhances the detection accuracy by 1–2 % by adaptively selecting only low-light images for enhancement instead of enhancing all images from varying light conditions. © 2025 Elsevier Ltd