Improved Steel Bar Detection Algorithm Based on Yolov8

Rebar Steel bars play an indispensable role in construction projects. It is very time-consuming to count a large number of steel bars at the construction site manually. Object detection methods can be used to count steel bars, but due to the dense end faces of steel bars, the boundary adhesion of steel bar end faces, the fusion of steel bar end faces and background, and uneven lighting, steel bar recognition is difficult. To address the above problems, this paper proposes an improved lightweight yolov8 model, replaces the SPPF module with the scale feature interaction AIFI module, improves the original PAFPN, and proposes Small Object Enhance Pyramid. Compared with the traditional method of adding a P2 detection layer, we use the P2 feature layer to obtain features rich in small target information through SPDConv and hand them over to P3 for fusion. Then, drawing on the idea of CSP and combining the OmniKernel based on AAAI2024, the CSP-OmniKernel is improved for feature fusion. The OmniKernel module consists of three branches: global branch, large branch, and local branch. It effectively learns feature representation from global to local, and finally improves the detection performance of small targets, with mAP@ 0.5 increased from 97.5 % to 98.5 %. Moreover, the number of parameters only increased by 0.92 MB, and the computing power increased from 8.2GFLOPs to 9.0GFLOPs, achieving high-precision deployment while maintaining low computing requirements. It is suitable for steel bar identification and counting at construction sites.