Fig. 1. Architecture of CNN.

Proposed CNN architecture consists of six convolutional layers, each with 3 × 3 kernel with 64 filter banks. To maintain original resolution throughout network, we excluded pooling layer in CNN structure and used original matrix size of 512 × 512 for both input and output images. Rectified linear unit was used as activation function at end of each convolutional layer. As we applied concept of residuals, proposed CNN was trained to learn difference between target and input images, and final image was obtained by adding residual image (output) to input. CNN = convolutional neural network