Traffic Sign Recognition Using Multi-Scale Enhanced Residual Network and Deformable YOLO in Intelligent Transportation

Abstract

With the continuous development of intelligent transportation systems, traffic sign recognition (TSR) under complex scenarios such as low resolution and poor lighting has become a critical research focus. This study proposes a two-stage TSR framework that combines a Multi-scale Enhanced Channel Residual Network (MECRN) for image enhancement and a Deformable Convolutional YOLO-based detection network (PP-YOLO-DCN). In the enhancement stage, MECRN integrates dense connections, multi-scale convolution, and channel attention mechanisms to improve image clarity and detail preservation. In the recognition stage, deformable convolutions and depthwise separable convolutions are introduced into the PP-YOLO framework to enhance the detection of small, irregular traffic signs while reducing computational complexity. Experimental results show that the MECRN achieves a peak signal-to-noise ratio (PSNR) of 31.7 dB and a structural similarity index (SSIM) of 0.897. In low-light scenarios, the learned perceptual image patch similarity (LPIPS) reaches 0.185, indicating superior visual restoration. The PP-YOLO-DCN model attains a mean average precision (mAP@0.5) of 0.91 and 0.86 under dense multi-target and adverse weather conditions, respectively, with real-time performance of over 40 FPS. Compared with baseline methods, the proposed framework significantly improves recognition accuracy and robustness in challenging traffic environments, providing effective technical support for intelligent transportation applications.