基于锐度感知最小化与多色域双级融合的视网膜图片质量分级

针对视网膜图片质量差异性大，质量分级模型泛化性能不足的问题，本文提出一种基于锐度感知最小化(Sharpness Aware Minimization，SAM)的多色域双级融合分级方法，用于视网膜图片质量评估。该方法首先采用RGB、HSV和LAB三种色域空间作为模型的特征提取空间，并利用ResNeSt作为特征提取网络，提取不同色域的空间特征。然后利用模型的特征级与预测级进行双级融合，再利用SAM优化方法提高视网膜图片质量分级模型的泛化性能。最后在EyeQ数据集上进行实验仿真，其准确率为87.35%、精确度为85.87%、敏感度为85.07%、F值为85.44%，所提方法的整体分级性能优于其他方法。

Retinal image quality grading based on SAM multi-color gamut and two-stage fusion

In view of the large difference in the quality of retinal images and the insufficient generalization performance of the quality grading model, this paper proposes a multi-color gamut two-level fusion grading method based on Sharpness Aware Minimization (SAM) for retinal image quality evaluation. The method first uses RGB, HSV, and LAB three-color gamut spaces as the feature extraction space of the model, and uses ResNeSt as the feature extraction network to extract spatial features of different color gamuts. Then the feature level and prediction level of the model are used for two-level fusion, and then the SAM optimization method is used to improve the generalization performance of the retinal image quality classification model. Finally, the experimental simulation is performed on the EyeQ data set. The accuracy is 87.35%, the accuracy is 85.87%, the sensitivity is 85.07%, and the F value is 85.44%. The overall classification performance of the proposed method is better than other methods.