To solve the problems that the existing detection algorithm of HRNetV2p could not balance the detection accuracy of defects at various scales, the feature fusion module combined with channel attention mechanism was introduced into the detection algorithm of HRNetV2p, which could adaptively adjust the ratio of spatial-semantic information in the fusion features and could improve the network ability to preserve semantic information in shallow features. The surface defect detection dataset of IIDD for malleable iron was established for data labelling and data statistics. The CG dense skip transmission unit and the CG adaptive fusion module were introduced into the HRNetV2p network for adaptively adjusting the spatial-semantic information ratio of the front-layer features through three operations of integration, recalibration and reintegration. The experimental setup and evaluation index were given, and the performance experiments of the improved HRNetV2p algorithm on the malleable iron surface defect dataset of IIDD were completed. The results show that the average detection accuracy AP50 of the improved HRNetV2p algorithm on IIDD is 91.3%, which is 2.6% higher than the average detection accuracy of the original HRNetV2p. The detection accuracies of large, medium and small scale defects are improved by 2.7%, 2.7% and 5.6%, respectively.
REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017,39(6):1137-1149.
［2］
REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]∥Proceedings of the 29th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway：IEEE Com-puter Society，2016:779-788.
DONG Y P, GAO C Q, CHEN F, et al. Infrared small target detection method based on attention mechanism［J］. Journal of Chongqing University of Posts and Telecommunications （Natural Science Edition）,2023,35(2)：219-226．（in Chinese）
LIU H H, SUN C, HE H Q, et al. Metal surface defect detection method YOLOv3I[J]. Journal of Jilin University (Science Edition), 2023,61(3):612-622.(in Chinese)
［5］
SUN K, ZHAO Y, JIANG B R, et al. High-resolution representations for labeling pixels and regions[J]. arXiv preprint arXiv:1904.04514v1.
［6］
SUN K, XIAO B, LIU D, et al. Deep high-resolution representation learning for human pose estimation[C]∥ Proceedings of the 2019 IEEE/CVF Conference on Com-puter Vision and Pattern Recognition. Piscataway：IEEEComputer Society，2019:5686-5696.
［7］
XU J J, SUN X, ZHANG Z Y, et al. Understanding and improving layer normalization[C]∥Proceedings of the 33rd Annual Conference on Neural Information Proces-sing Systems. [S.l.]: Neural Information Processing Systems Foundation,2020:4358-4368.
［8］
MISRA D. Mish: a self regularized non-monotonic activation function[J]. arXiv preprint arXiv:1908.08681,2019.
［9］
BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4: optimal speed and accuracy of object detection[J]. arXiv preprint arXiv:2004.10934v1.
［10］
LI Y H, CHEN Y T, WANG N Y, et al. Scale-aware trident networks for object detection[J]. arXiv preprint arXiv:1901.01892.
