Steel surface defect detection based on YOLOv8-SOE

In industrial applications such as steel surface defect detection, small target detection remains a challenging task due to the limited resolution of conventional detection layers, making it difficult to capture fine-grained defect details. Although YOLOv8 has shown remarkable performance in multi-scale target detection and complex environments, the model struggles with small target detection, especially when it comes to tiny defects on steel surfaces. Traditional solutions to enhance small target perception, such as adding additional detection heads like the P2 layer, often lead to increased computational overhead and inference time. To address this issue, this paper proposes an improved YOLOv8-SOE model that specifically enhances the detection performance for small defects on steel surfaces. The YOLOv8-SOE model incorporates several innovations aimed at improving both detection accuracy and computational efficiency. First, a novel feature processing module, FSCConv, is introduced to handle the P2 layer features. FSCConv leverages dilated convolutions to capture multi-scale contextual information while preserving fine details of small targets. This approach enhances small target perception without the need for additional detection layers, thus avoiding the computational burden typically associated with such modifications. Next, the processed P2 features are fused with P3 layer features, improving small target detection further without incurring significant computational costs. To optimize the feature fusion process, a cross-stage local network combined with Omni-Kernel (CSP-OK) is proposed. CSP-OK primarily leverages the CSPNet approach to reduce redundant gradient computations and integrates the Omni-Kernel to prevent the repetitive extraction of similar information at different layers, thereby improving information utilization efficiency. This optimization reduces redundant information computations and effectively utilizes inter-layer features, resulting in a more efficient and detailed integration of multi-scale information. In addition, the model uses the SIoU loss function for bounding box regression. This loss function not only considers the overlap between the predicted and ground truth boxes but also incorporates their distance, angular deviation, and shape similarity. By integrating these factors, the SIoU loss function provides a more comprehensive optimization strategy, thereby improving the accuracy of target localization. Experimental results demonstrate that the YOLOv8-SOE model achieves a mean average precision (mAP) of 80.7% on the NEU-DET dataset, a 5.4% improvement over the baseline YOLOv8 model. The model also exhibits excellent generalization ability on the VOC2012 dataset. Overall, the proposed YOLOv8-SOE model significantly enhances small target detection precision while maintaining high computational efficiency, making it a promising solution for real-world industrial defect detection applications.