融合离散小波变换的YOLOv11晶圆缺陷检测算法

Wafer Defect Detection Algorithm Based on YOLOv11 with Wavelet Transform Integration

摘要: 针对半导体制造过程中晶圆表面微小缺陷尺度小、特征弱且易受复杂背景干扰，导致传统目标检测方法存在漏检率高、定位精度不足等问题，本文提出了一种融合离散小波变换(Discrete Wavelet Transform, DWT)特征保留机制的YOLOv11晶圆缺陷检测算法。该方法在主干网络下采样阶段引入WaveletConv模块，以二维离散小波变换替代传统步长卷积，实现特征图的多频分解与结构化表征。具体而言，输入特征被分解为低频全局分量以及水平、垂直和对角方向的高频细节分量，在完成空间降采样的同时，有效保留了与微小缺陷识别密切相关的边缘、纹理和局部突变信息。进一步地，通过通道维融合策略将多频信息联合传递至深层网络，增强了模型对微弱缺陷特征的表征能力与判别能力，从而缓解了传统卷积下采样过程中高频信息衰减带来的性能损失。实验结果表明，所提出的方法在晶圆缺陷数据集上取得了优于基线模型的检测性能，在微小缺陷检出率和定位精度方面表现出明显优势。研究表明，将小波域多尺度、多频特征建模引入目标检测框架，能够有效提升模型对复杂工业场景中微观缺陷的感知能力，为半导体制造过程中的高精度视觉检测提供了新的解决思路。

Abstract: In semiconductor manufacturing, wafer surface defects are characterized by small scales, weak features, and susceptibility to complex background interference, which leads to high miss-detection rates and insufficient localization accuracy in traditional object detection methods. To address these challenges, this paper proposes a YOLOv11-based wafer defect detection algorithm integrated with a Discrete Wavelet Transform (DWT) feature preservation mechanism. This method introduces a WaveletConv module at the downsampling stages of the backbone network, replacing conventional strided convolutions with two-dimensional discrete wavelet transform to achieve multi-frequency decomposition and structured representation of feature maps. Specifically, the input features are decomposed into low-frequency global components and high-frequency detail components in horizontal, vertical, and diagonal orientations. This approach accomplishes spatial downsampling while effectively preserving edge, texture, and local mutation information crucial for tiny defect recognition. Furthermore, through a channel-wise fusion strategy, the multi-frequency information is jointly propagated to deeper network layers, enhancing the model’s representation and discriminative capabilities for subtle defect features, thereby mitigating performance degradation caused by high-frequency information attenuation during traditional convolutional downsampling. Experimental results demonstrate that the proposed method achieves superior detection performance compared to baseline models on wafer defect datasets, exhibiting significant advantages in tiny defect detection rates and localization precision. The study reveals that introducing wavelet-domain multi-scale, multi-frequency feature modeling into object detection frameworks can effectively improve the model’s perception capability for microscopic defects in complex industrial scenarios, providing novel insights for high-precision visual inspection in semiconductor manufacturing processes.