关键词: 机器视觉, 微缺陷检测, 边缘信息增强, 上下文融合扩散金字塔, YOLOv11

Abstract: To address challenges such as small target size, significant scale variation, complex backgrounds, and low detection accuracy in micro-defect detection on wafer mold surfaces, this paper proposes the MMF-YOLO algorithm, which combines edge information focusing and context information fusion diffusion. Firstly, the edge information focusing module (EIFM) is introduced to enhance the C3k2 module in the original network, enabling the selection of key features highly correlated with the target from multi-scale edge information. Secondly, the context-fusion diffusion pyramid network (CFD-PN) structure is employed to optimize the neck network, reducing feature confusion and loss during information fusion by extracting multi-level representations of features in terms of spatial resolution and semantic information. Additionally, the adaptive down-sampling module (ADown) is incorporated to optimize the number of parameters and computational redundancy in convolutional layers, thereby reducing model complexity. Finally, the feature scale-aware detection head (FSDH) is utilized to minimize network storage overhead by employing shared convolutions. Experimental results demonstrate that the MMF-YOLO algorithm achieves a 6.93 percentage points improvement in mAP@0.5 compared to the baseline YOLOv11n on the wafer mold surface micro-defect dataset, making it more suitable for micro-defect detection tasks on wafer mold surfaces and deployment on embedded platforms for efficient inference.

Key words: machine vision, micro-defect detection, edge information enhancement, context-fusion diffusion pyramid, YOLOv11