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

Retinal Image Quality Grading Based on Sharpness Perception Minimization and Multi-gamut Bi-level Fusion

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,提出了一种基于锐度感知最小化的多色域双级融合算法,用于视网膜图片质量的分级预测。首先,采用ResNeSt网络对RGB(red, green, blue)、HSV(色相hue、饱和度saturation、亮度value)和LAB(L表示像素的亮度、A表示从红色到绿色的范围、B表示从黄色到蓝色的范围)3种色域空间进行特征提取。其次,使用网络的特征输出与预测输出进行双级融合,丰富视网膜图片的特征表示。然后,使用锐度感知最小化对视网膜图片质量分级模型进行优化,提高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%,所提算法能有效区分视网膜图片的质量等级并提高模型的泛化性能。

In view of the large difference in retinal image quality and the insufficient generalization performance of the quality classification model,a multi-color gamut bi-level fusion algorithm based on sharpness perception minimization was proposed for the graded prediction of retinal image quality.Firstly,the ResNeSt network was used to extract features from three color gamut spaces,RGB(red,green,blue),HSV(hue,saturation,value)and LAB(L represents the brightness of the pixel,A represents the range from red to green,and B represents the range from yellow to blue)Second,the bi-level fusion was composed of feature output and prediction output of the network and enriched the feature representation of retinal images.Then,the retinal image quality grading model was optimized using sharpness-aware minimization to improve its generalization performance.Finally,the experimental simulation was carried out on the EyeQ dataset,and the accuracy rate is 87.35%,the precision is 85.87%,the sensitivity is 85.07%,and the F value is 85.44%.The quality level of retinal images can be effectively distinguished by the proposed algorithm,and the generalization performance of the model can also be improved.