Introduction: This network has been implimentd by following the basic idea of UNet. A Residual block from ResNet and Squeeze and Excitation block from SENet has been used. Residual block allows training deeper networks and prevent from vanishing gradient problem Whereas, in Squeeze block global information of the feature map is extracted through its spatial dimension then provided as input to the Excitation block. This block extract the channel wise information and corelation between them. Therefore emphasizing on most relevant features and suppress less important one. A hybrid deep convolution neural network (RSE-UNet) is shown below.

Libraries Used

Pytorch, Numpy, Matplotlib, CV2, Tensorboard, Tmux, Sk-learn