Multi-level Fused Lightweight YOLOv7 for Traffic Sign Detection
Jingyan Wang, Aiqing Huo, Haowei Zhang · 2025
To address the challenges associated with the large number of parameters in the YOLOv7 network, its low detection accuracy, and suboptimal real-time performance during traffic sign detection, a multi-level fused lightweight SSPI-YOLOv7 (ShuffleNetv2 SimAM PConv Inner-CIoU-YOLOv7) model is proposed. Initially, the backbone network of the YOLOv7 model is substituted with a light-weight Shuf-fleNetv2 network, aiming to decrease both the network parameters and computational complexity. Initially, the backbone network of the YOLOv7 model is substituted with a lightweight Shuf-fleNetv2 network, aiming to decrease both the network parameters and computational complexity. SPPCSPC modules, further reducing network parameters and computational demands. Additionally, by reconstructing the SPPCSPC module and incorporating the parameter-free SimAM Additionally, by reconstructing the SPPCSPC mod-ule and incorporating the parameter-free SimAM attention mechanism, not only are the network parameters minimized, but attention to shallow targets is also Additionally, by reconstructing the SPPCSPC module and incorporating the parameter-free SimAM attention mechanism, not only are the network parameters minimized, but attention to shallow targets is also heightened, thereby enhancing detection accuracy. Lastly, the network's loss function CIoU is refined to Inner-CIoU, which bolsters target position information extraction without altering the model's parameters, leading to improved detection accuracy. of these enhancements: post-improvement, the parameter count dropped from 37.2M to 16.3M, a reduction of 56.2%; the weight file size decreased from 71.3 M to 30.7M; the frames per second (FPS) increased from 83 to 110; GFlops were reduced from 105.1 to 28.4; and the mean average precision (mAP) achieved a score of 96.6%.