Computer Vision News - February 2023

32 Spectral Imaging Tool The default config file looks like this: # Default image super-resolution configuration Task _ Options: task: 'Super-Resolution' # deep learning task [Calibration/ Classification/Denoising/Segmentation/Super-Resolution] classes: 0 # no. classes for Classification or Segmentation, zero otherwise Training _ Options: training _ option: 'Train From Scratch' # training option ['Apply Pre-Trained Network'/'Train From Scratch'/'Transfer Learning'] pretrained _ network: 'None' # path to pretrained network, 'None' if training _ option = 'Train From Scratch' pretrained _ classes: 0 # no. classes for pretrained Classification or Segmentation model, zero otherwise Network _ Hyperparameters: network: RCAN # network architecture [UNet/ ResUNet/ResNet/RCAN] dimension: '2D' # convolution dimension for images ['2D'/'3D'] activation: ReLU # activation function [ReLU/ LeakyReLU/PReLU] normalization: 'None' # normalization layer [None/ BatchNorm/LayerNorm/InstanceNorm/GroupNorm] Training _ Hyperparameters: epochs: 1 # no. training epochs batch _ size: 2 # batch size learning _ rate: 0.0001 # learning rate input _ image _ size: 16 # input image size (only applies to Super-Resolution) target _ image _ size: 64 # target image size (only applies to images) spectrum _ length: 500 # spectrum length optimizer: Adam # optimizer [Adam/Adagrad/SGD/ RMSprop] scheduler: Constant # learning rate scheduler [Constant/Step/Multiplicative/Cyclic/OneCycle/ReduceOnPlateau]