PHG-YOLO: A Lightweight Steel Surface Defect Detection Algorithm based on YOLOv8

The existing deep learning-based steel surface defe ct detection algorithm has the problems of low detection accura cy and large number of parameters, and the limited computing power of some enterprises leads to the difficult deployment of d etection algorithms. To address these issues, this paper propose s a lightweight improved algorithm called PHG-YOLO. First, a new feature extraction module, C2f-PKI, is introduced into the backbone network to replace the original C2f module, utilizing non-expanded multi-scale convolutional kernels to better extra ct target features at different scales. Second, an High-level Scree ning-feature Pyramid Networks(HSFPN) is introduced in the n eck network, using high-level features as weights to filter low-le vel feature information through a channel attention module, an d then merging the filtered information with high-level features to enhance the model's feature representation capability. Finally, we propose a lightweight parameter-shared detection head (G N-Head), which effectively improves detection accuracy while maintaining a lightweight design. Extensive experiments on the NEU-DET dataset show that the model's parameter count and computational complexity are 5.08x10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</sup>and 19.8 GFLOPs, only 45.6% and 68.9% of YOLOv8s, respectively. With this lightwei ght approach,the mAP50 reaches 79.0%, an improvement of 4. 4% over YOLOv8s.