Abstract: During the hot rolling process, the precise positioning of steel billets is of paramount importance, as it enables automated billet inspection, which in turn enhances both production efficiency and quality control. However, the actual production process is arduous due to the harsh conditions of the billet production environment, which presents challenges for plate inspection. To address the problems mentioned above, a robust Yolov8 algorithm for billet detection is proposed. A hybrid attention mechanism network is proposed to solve the problem of shallow feature loss in pyramid networks, enhancing the model's ability to learn local features and improving detection accuracy while maintaining model lightweightness. The implementation method involves the introduction of attention modules into the network with the objective of enhancing the preservation of detailed image feature information, thereby improving the overall detection accuracy of the targets. Subsequently, the NPANet feature fusion structure is designed to enhance the network's ability to fuse multi-scale features of images. This is achieved by refining the convolution modules to make the network model lighter. Finally, the loss function is improved to enhance the algorithm's regression performance and reduce the error in stable box generation. The experimental results demonstrate that the improved NDS-yolov8 model, in comparison to the initial network structure, exhibits a reduction in the weight file size from 6.2 MB to 4.6 MB, a decrease infloating-point performance from 8.1 GFLOPS to 6.4 GFLOPS, and an increase in the mean average precision (P_mA@0.5:0.95by 0.5% at different I_oU values. Compared to actual values in real-world scenarios, the NDS-yolov8 network model demonstrates significantly reduced error margins relative to the original Yolov8 network model. It achieves more accurate estimation of the billet's real-time position, thereby effectively enhancing the performance of billet detection and localization.