Using Transfer Learning and BPDFHE to Improve Ocular Image Recognition Accuracy

Abstract

We used image enhancement algorithms along with transfer learning to fine-tune a deep convolutional neural network to perform ocular image recognition. To enhance the input images, we used a novel color image histogram equalization technique called Brightness Preserving Dynamic Fuzzy Histogram Equalization, which showed significant accuracy improvements: on the test data, using AlexNet, the ROC Area Under the Curve (AUC) increased to over 0.99, Equal Error Rate (EER) decreased 4-fold and dropped below 4%, and decidability (a measure of class separability) increased from 1.89 to 4.17