目的 针对玻璃的材料透明性以及条带噪声等固有属性使得传统玻璃缺陷分割算法准确率较低等问题，提出一种基于双特征高斯混合模型的玻璃缺陷分割方法。方法 首先，利用分数阶运算对玻璃缺陷增强，用灰度共生矩阵获取纹理特征，从而构建玻璃缺陷的双特征向量;将双特征向量引入高斯混合模型，并利用马尔科夫随机场的相邻像素空间信息对玻璃缺陷分割高斯混合模型进行改进，通过交替进行玻璃缺陷像素点与标号场之间映射关系的估计和基于高斯核函数空间约束更新，完成玻璃缺陷分割;最后，应用模糊熵对缺陷图像分割结果进行后续处理。结果 对疖瘤、污点、气泡以及夹杂等4种典型缺陷样本图像进行性能测试和不同算法对比分析实验，实验结果表明，所提算法的 指标达到98.59%， 指标达到7.03%，衡量指标优于其他算法。结论 将灰度特征和纹理特征引入玻璃缺陷分割的马尔科夫随机场，能够抑制非缺陷目标，并保留低对比度玻璃缺陷，提高玻璃缺陷分割算法的鲁棒性和准确性。

The work aims to propose a glass defect segmentation method based on a dual feature Gaussian mixture model to solve the low accuracy of traditional glass defect segmentation algorithm caused by inherent properties of glass material, such as transparency and stripe noise. Firstly, fractional calculation and gray-level co-occurrence matrix were adopted to enhance glass defects, and obtain texture features, respectively, and thereby construct dual feature vectors of glass defects. The dual feature vector was introduced into the Gaussian mixture model, and the adjacent pixel spatial information of the Markov random field was used to improve the glass defect segmentation Gaussian mixture model. Then, the glass defect segmentation was completed by alternately performing the estimation of the mapping relationship between the glass defect pixels and the label field and the updating based on the space constraint of the Gaussian kernel function. Finally, fuzzy entropy was applied to the subsequent processing of the defect image segmentation results. The performance test and comparative analysis experiment of different algorithms were performed on four typical defect sample images of furuncle, stain, bubble and inclusion. The experimental results showed that the Dice index of the proposed algorithm reached 98.59% and the Mcr index reached 7.03%, which was better than that of other algorithms. Introducing gray-level features and texture features into the Markov random field of glass defect segmentation can suppress non-defect targets, retain low-contrast glass defects, and improve the robustness and accuracy of the glass defect segmentation algorithm.