Granularity Classification and Feature Fusion Methods in Traffic Sign Detection

Abstract

Nowadays, deep learning based on detection algorithms have replaced the traditional manual feature extraction target algorithms and have achieved amazing results in many places with their powerful automatic feature extraction capabilities. However, the results are not ideal for the detection of small targets with low resolution and a lot of noise, such as traffic signs. To address the current problems of slow detection speed and low detection accuracy in small target detection, this paper adopts a feedback-driven mechanism to solve the image level imbalance of the input feature space under the original data distribution. At the same time, this paper designs a novel and flexible two-stage traffic sign recognition framework. The complex task of traffic sign detection and recognition is decomposed into two stages: 1) designing a superclass classifier to more accurately separate traffic signs in complex natural scene images; 2) The idea of similarity metric learning is used to design fine-grained classifiers to recognize traffic signs. Finally, to verify the effectiveness of the model, the model was first compared with Faster R-CNN and found to possess higher detection accuracy; then the model was experimented with the R-FCN model on TTIOOK dataset and CCTSDB dataset respectively, and the comparison of the experimental results revealed that the model improved over the R-FCN model in most of the metrics in the traffic sign detection task.