Research on Lightweight Detection Network for Fog Cannon Vehicles Based on Multi Scale Feature Fusion

In response to multiple incidents of fog spray operations interfering with gas station detection results near air monitoring stations in China, this study collected and supplemented fog spray vehicle datasets, and proposed a lightweight detection network for fog spray vehicles based on multi-scale fusion on the basis of the YOLOv8 network model structure. Firstly, in the backbone section, AKConv and EIEM modules are used instead of C2f, which give convolution kernels any number of parameters and arbitrary sampling forms, providing rich choices for balancing network overhead and performance; Secondly, the GL BiFPN structure is fused in the head section to reduce incorrect information in advanced features, enhance feature fusion capabilities, and reduce network computational complexity; Afterwards, WI-ShapeIoU was used as the loss function to improve the model, making the bounding box regression more accurate and enhancing the detection performance and model universality. The experimental results show that, compared with YOLOv8s network, the integrated detection accuracy of fog artillery spray is improved by 4.1%, the floating point computation is reduced by 14.09%, and the FPS is improved to 142.3 frames/second. Compared with the current mainstream detection models, the improved network reduces parameters and improves detection speed while maintaining improved accuracy.