Age Estimation from Retinal Images: Different Image Preprocessing Approaches

Abstract

Human age is considered an important biometricparameter that is often difficult to determine. Previous studieshave shown that the non-specific general anatomical and physiologicalcharacteristics seen on fundus images are all likely signs ofageing. This paper focuses on age estimation from retinal imageswith different image preprocessing approaches together withproposed image detail enhancement method. Convolution neuralnetwork framework is based on the ResNet-34 architecturetogether with the Consistent Rank Logits algorithm estimatingage as an ordinal variable. The best model achieved a meanabsolute error of 3.47 years, outperforming existing modelsestimating age from retinal images.