Efficient feature fusion network for small objects detection of traffic signs based on cross-dimensional and dual-domain information

Abstract The objectives in traffic sign detection and recognition(TSDR) scenario are predominantly small, which frequently result in missed and erroneous detection due to their limited information content and complex environment. To address these problems, this paper proposes a new network architecture CDFF-YOLO(Cross-dimensional and Dual-domain Feature Fusion-You Only Look Once) which is integrated of various modules. For the purpose of overcoming the difficulty in extracting information from small objects, the MSF(Multi-dimension Spatial information Fusion) module in the network are used to extract feature sequences at different dimensions by superposition. So as to address the issue of the loss of detail information of small objects, embed the MPA(Multi-branch Perceptual Attention) module into the C2f module to capture feature information and enhance the global-local feature information exchange. In order to solve the issue of uneven illumination and occlusion in the detection scene. The DFF(Dual-domain Feature Fusion) module is employed to transform and fuse the extracted small object feature information from the SPDConv(Space-to-depth convolution) at the frequency and spatial domains, thereby enhancing the network's capability to reconstruct and fuse feature information in dual-domains. The experimental data on the TT100K dataset demonstrate that the enhanced algorithm exhibits an increase of 3.7\% in mAP@50, 4.8\% in mAP@50:95, and a 4.5\% and 3.7\% rise in the AP(average precision) and AR(average recall) for small objects, respectively. Additionally, the FPS(Frames Per Second) remains 157. The improved algorithm also performs well on the CCTSDB dataset. It is evident that the CDFF-YOLO algorithm has the capacity to markedly enhance the detection efficacy of traffic signs, while maintaining optimal detection speed.