Fig. 6. Neural network architecture.

Detailed schematic of the CNN used for training and testing. IB (Input Block), OB (Output Block), Bi (Block i, where i = 1, 2,…, 10), Pool (Max-pooling), Reshape (reshaping unit). The input block maps the input images via 64 convolutional filters. Each middle block (B1–B10) contains two convolution layers followed by a reshape and max-pooling layer, which together downsample the widths and heights of the images by a factor of two. A rectified linear unit (Relu) transform is placed after each convolution unit in the hidden layers. The images are then mapped to the output channel via the convolution filters in the output block. The MSE between the labels and the processed images is then calculated and back propagated to the network to update the learnable variables