Improved Lightweight YOLOv8 With DSConv and Reparameterization for Continuous Casting Slab Detection on Embedded Device

The molten steel is transformed into slabs in the process of continuous casting. These slabs are subsequently conveyed to the hot rolling area for further processing or stored in the slab storage area for future utilization. An improved lightweight object detection algorithm called You Only Look Once version 8 (YOLOv8) with DSConv, light Efficient RepGFPN, and CARAFE (DSCLERC-YOLOv8) was proposed in this article for real-time detection of slabs transmitted on the roller table, to be applied in embedded devices. DSCLERC-YOLOv8 utilizes an enhanced C2f with distribution shifting convolution (C2f_DSConv) as the feature extraction module, which enhances the processing speed of the object detection algorithm while maintaining accuracy. Moreover, the algorithm incorporates an improved Lightweight Efficient-reparameterization generalized feature pyramid network (RepGFPN) and content-aware reassembly of features (CARAFE) for feature map upsampling, enhancing the accuracy and efficiency of image prediction. In addition, the article investigates the impact of reducing network stages on the algorithm. For practical applications, a P4 network structure is employed, resulting in a slab detection accuracy rate of 86.91% for mAP50:95. Although this accuracy is slightly lower than YOLOv8n (87.78%) by 0.87%, the significant improvement of this algorithm lies in a 14.7% reduction in network latency, a 77.22% reduction in parameter count, a 66.29% reduction in computational complexity, and a 62.90% reduction in model size. Finally, by implementing model pruning, quantization, and deploying with the TensorRT engine, the algorithm successfully achieves a real-time speed of 76 FPS on Jeston Orin NX. This algorithm offers a dependable solution for the real-time detection of steel slabs during continuous casting and rolling processes in the steel industry.