AIF-YOLO: small traffic sign recognition based on YOLOv7

Abstract

Traffic sign recognition is an essential part of intelligent transportation systems and autonomous driving. However, due to the small size of traffic signs, there are a few problems in traffic sign recognition, such as missed recognition, false recognition, and so on. Therefore, we choose YOLOv7 as the baseline and propose an anti-interference feature enhancement YOLO, abbreviated as AIF-YOLO. Firstly, to enhance the network’s feature extraction capability, we add an adaptive feature extraction module (AFEM) structure to adaptive adjust the size of receptive field and refine features. Secondly, we introduce a contextual transformer into the feature extraction network of YOLOv7, resulting in the contextual transformer efficient layer aggregation network (CoT-ELAN) module. This improvement enables the model to capture the relation between traffic signs and environment and thus separate the interference signs. Furthermore, Wise-IoU loss was adopted to optimize network training, effectively addressing the poor quality label issue in the traffic sign dataset. Experiments were tested on Tsinghua-Tencent 100k dataset(TT100K), and the results show that our improved model performed better than the baseline YOLOv7, and other popular algorithms on precision, recall, and mAP@0.5.