Traffic sign recognition algorithm based on feature pyramid attention

Abstract

The traditional traffic sign recognition algorithm is easily affected by factors such as complex background and illumination in a real scene, which can easily lead to missed detection and misdetection, in the stage of traffic sign detection, the Feature Pyramid Spatial Attention (FPSA) module is proposed. In the process of generating the feature pyramid, the FPSA module uses high-level features as the attention information of low-level features. In order to solve the problem of sample imbalance in the dataset, the dataset is screened, and the category loss weight based on the effective number of samples is introduced when calculating the loss. Experiments conducted on 45 traffic sign categories on the TT-100K dataset prove that FPSA and the proposed category loss weights are beneficial to improve the performance of small object detection and recognition in complex backgrounds, and are obtained on the Faster R-CNN model 78.9% mAP accuracy.