Abstract:

To address the challenge of detecting small and low-contrast defects in complex microscopic images, a defect method technology based on hierarchical microscopic feature modeling and simulation is proposed. The method is built on the RT-DETR (real-time detection transformer)framework to construct the HM-RTDETR (hierarchical microscopic RT-DETR) model. It maintains the global feature modeling ability of the Transformer and introduces a Dense O2O-Mosaic,a high-density one-to-one Mosaic augmentation strategy, to increase supervision density for small samples. A depthwise separable convolution (DWConv) module is used to enhance local detail extraction in microscopic textures, and a learnable PatchExpand module is applied for spatial semantic reconstruction, improving the identification of tiny defects. Furthermore, thecooperative fusion of DWConv and PatchExpand achieves adaptive multi-scale feature integration and lightweight optimization.Experimental results show that the proposed model maintains consistent advantages under both offline augmentation settings: HM-RTDETR achieves mAP0.5-0.95 of 57.8% and 70.4% under Setting I and Setting II, respectively, improving over the corresponding baselines by 18.9% and 21.6%.The method achieves a good balance between accuracy and efficiency, providing an effective and extensible solution for automatic microscopic defect detection and simulation analysis.

Key words: microscopic defect detection, RT-DETR(real-time detection transformer), Mosaic data augmentation, DWConv, PatchExpand, lightweight network