Open Access

ARTICLE

Pre-Locator Incorporating Swin-Transformer Refined Classifier for Traffic Sign Recognition

by Qiang Luo¹, Wenbin Zheng^1,2,*

1 School of Software Engineering, Chengdu University of Information Technology, Chengdu, 610225, China
2 V.C. & V.R. Key Lab of Sichuan Province, Sichuan Normal University, Chengdu, 610068, China

* Corresponding Author: Wenbin Zheng. Email:

(This article belongs to the Special Issue: Computer Vision and Machine Learning for Real-Time Applications)

Intelligent Automation & Soft Computing 2023, 37(2), 2227-2246. https://doi.org/10.32604/iasc.2023.040195

Received 08 March 2023; Accepted 10 May 2023; Issue published 21 June 2023

Abstract

In the field of traffic sign recognition, traffic signs usually occupy very small areas in the input image. Most object detection algorithms directly reduce the original image to a specific size for the input model during the detection process, which leads to the loss of small object information. Additionally, classification tasks are more sensitive to information loss than localization tasks. This paper proposes a novel traffic sign recognition approach, in which a lightweight pre-locator network and a refined classification network are incorporated. The pre-locator network locates the sub-regions of the traffic signs from the original image, and the refined classification network performs the refinement recognition task in the sub-regions. Moreover, an innovative module (named SPP-ST) is proposed, which combines the Spatial Pyramid Pool module (SPP) and the Swin-Transformer module as a new feature extractor to learn the special spatial information of traffic sign effectively. Experimental results show that the proposed method is superior to the state-of-the-art methods (82.1 mAP achieved on 218 categories in the TT100k dataset, an improvement of 19.7 percentage points compared to the previous method). Moreover, both the result analysis and the output visualizations further demonstrate the effectiveness of our proposed method. The source code and datasets of this work are available at .

---

Keywords

---