Variants of ReLU in 2-Layer Sequential CNN to Extract Hard Exudates in Diabetic Retinopathy

Rectified linear units are used conventionally as activation function in deep learning to modulate the inputs at the intermediate stages of convolutional neural networks. Pooling functions are used later to the ReLU in order to perform the predictive and descriptive operations on the input data. There are parametric and nonparametric, static and dynamic categories of ReLU. Various input samples are tested to assess the performance of CNN using different variants of ReLU. Digital fundus images of diabetic retinopathy are color restricted and represent various types of biological signs.