Abstract
Steel is a crucial raw material in the industrial sector, and its surface defects significantly impact product quality. These defects are diverse in type, complex in shape, uneven in distribution, and varied in size, posing substantial challenges for performance and detection. Addressing the limitations of current deep learning-based steel surface defect detection algorithms in feature extraction, feature fusion, and multi-scale defect recognition-leading to high false detection rates, frequent missed detections, and low detection accuracy-this paper proposes a steel surface defect detection algorithm based on an improved YOLOv8n, named SCCI-YOLO. Firstly, the SPD-Conv module is introduced into the backbone network, utilizing adaptive weight allocation to focus convolutional kernels more on critical regions of the image, thereby improving detection accuracy for low-resolution and small objects. Secondly, we designed the C2f_EMA module, aimed at extracting more useful feature information and enhancing feature fusion effects. To further enhance the detection capability for small defects, a lightweight cross-scale feature fusion module (CCFM) is incorporated into the Neck network. This module integrates features from different scales, enhancing the model's adaptability to scale variations and improving detection accuracy for small-scale objects. Finally, to address the weak generalization and slow convergence issues of existing IoU loss functions across different detection tasks, we employed the Inner-IoU loss function, which improves the model's convergence speed and regression accuracy.The experimental results show that SCCI-YOLO achieves a mAP of 78.6% on the NEU-DET dataset, which improves the detection accuracy by 2.2% and 5.9% compared to YOLOv8n and YOLOv7, respectively, and reduces the number of model parameters by 43.9% compared to the original model, and the study demonstrates that the algorithm exhibits an excellent overall performance in the detection of defects on steel surfaces.