LSKA-YOLOv8: A lightweight steel surface defect detection algorithm based on YOLOv8 improvement

In order to solve the problem of difficult deployment of existing deep learning-based defect detection models in terminal equipment with limited computational capacity, a lightweight steel surface defect testing model LSKA-YOLOv8 was proposed based on the YOLOv8n target detection framework. The model uses KernelWarehouse Conv (KWConv) with lower computational volumes, which significantly reduces the required computational resources. By updating the traditional Feature Pyramid Network (FPN) structure to Bi-directional Feature Pyramid Network (BiFPN), enhance the capture of contextual information and reduce the number of parameters of the model. In addition, the Spatial Pyramid Pooling Fast (SPPF) module was replaced with a more accurate Receptive Field Block (RFB) module, expanding the model’s sensory field and improving characteristic representation. At the same time, the Large Separable Kernel Attention (LSKAttention) module was introduced in the detection head, effectively enhancing the understanding and capture of the target characteristics, thus significantly improving the overall detection performance. Experiments on the NEU-DET dataset showed that the average accuracy of LSKA-YOLO increased by 4.4% on the map@0.5 indicator compared to the benchmark model, while the number of parameters and calculations of the model decreased by 26.7% and 50%, respectively. Provides valuable references and practical applications for deployment of defect detection models on computing-resource-limited terminal devices.