Normalization attention and lightweight convolution-based network for traffic sign detection

Abstract

Object detection and image recognition are considered to be the most critical and challenging aspects of computer vision. The rapid advancement of driverless technology and assisted driving systems further emphasizes the importance of traffic sign detection and recognition. Detecting small objects against complex backgrounds in practical assisted driving tasks remains a formidable challenge. Many existing target detection methods struggle to strike a balance between performance and parameters. In response, we propose a novel traffic sign detection model based on YOLOv8s-A Traffic Sign Detection Network incorporating Normalization-based Attention and Lightweight Convolution (YOLOv8-NL). The objective is to optimize the performance and address the challenge of balancing model parameters in traffic sign recognition within the transportation domain. Firstly, Normalization-based Attention (NAM) is introduced to address issues related to inconsistent input dimensions, uneven target distribution, and complex background of optimization. Secondly, lightweight Global Sparse Convolution (GSConv) is integrated to reduce the model parameters and enhance model generalization performance. Finally, to further enhance detection accuracy, the Wise-IOUv3 loss function is introduced to tackle difficulties associated with low-quality labeling of data. The effectiveness of the novel YOLOv8 model, which integrates Normalization-based Attention and Lightweight Convolution, is demonstrated through experiments conducted on two publicly available datasets. These experiments showcase the model's ability to significantly reduce parameters without compromising its guaranteed performance. It is noteworthy to note that significant improvements were observed in the CCTSDB2021 dataset, with a reduction of 1.2 million model parameters, resulting in 1% increase in mean average precision (mAP) compared to the YOLOv8s model. Furthermore, the mAP for the three-class target detection task on the TT00K dataset reached an impressive 94.2%.