轻量动态融合Transformer晶圆模具微缺陷检测(LDF-DETR)OA

To address the challenges of low signal-to-noise ratio,severe pseudo-defect interference,limited detection accuracy,and high industrial deployment complexity in the detection of micro-defects on wafer mold surfaces,this paper proposes a lightweight dynamic fusion detection model,LDF-DETR.This method is an enhancement of the RT-DETR architecture,where a dynamic alignment fusion(DAF)module is embedded in the backbone network to enable spatial semantic adapta-tion and parametric feature reorganization across multi-branch features.Innovatively,a learnable position encoding mecha-nism is introduced,which optimizes the potential spatial distribution of position embedding vectors through gradient feed-back,thereby improving the model's ability to perceive microscopic defect location features.To enhance computational efficiency,a dynamic group convolution with channel shuffle mechanism is proposed,effectively reducing computational redundancy by integrating channel reallocation.Experiments conducted on an industrial-grade micro-defect dataset for wafer molds demonstrate that LDF-DETR achieves a simultaneous reduction of 34.3 GFLOPs in computational cost and 8.38×106 parameters,while maintaining a mAP@0.5 detection accuracy of 93.72%,significantly outperforming existing light-weight object detection models.