SD-MERMF: A novel Swin-DeepLabv3 + network for gear surface defect detection based on mixture of experts routing mechanism and multilevel feature refinement fusion

To address issues of complex surface textures and backgrounds, multi-scale defects, and the limited ability of traditional networks to construct long-range dependencies and capture local representations of traditional network during the process of surface defect detection, a novel Swin-DeepLabv3 + network for gear surface defect detection based on a mixture of experts routing mechanism and multilevel feature refinement fusion (SD-MERMF) is proposed. Firstly, the SADME module is designed to break through the limitations of the single-induction bias, which not only dynamically suppresses background noise but also precisely fuses key defect features. And then, the proposed SSCFR module utilizes asymmetric convolution and cross-attention mechanisms to enhance the robustness of local details. Finally, the BRG module is employed to further improve the small defect boundary segmentation precision. Extensive comparative experiments and ablation experiments are carried out on self-built NEU-GSD dataset and two public NEU-SEG/IC-HQ datasets to verify the effectiveness and generalization ability of the proposed SD-MERMF model. The experimental results indicate that the proposed model achieves outstanding defect classification, segmentation, and detection tasks while achieving mIoU of 82.87 ± 0.14%, 85.97 ± 0.19% and 71.44 ± 0.28%. Furthermore, the proposed model avoids excessive computational costs and obtains a favorable balance between segmentation precision and inference burden against baseline approaches. Hence, the proposed SD-MERMF can be widely applied in gear maintenance or production processes for detecting gear surface defects.