Fig. 2. Overall architecture of the FlatNet.
The lensless camera measurement is first mapped into an intermediate image space using a trainable camera inversion layer. This stage is implemented separately for the separable and the non-separable case. A U-Net [31] then enhances the perceptual quality of the intermediate reconstruction. We use a weighted combination of three losses in training our network: a perceptual loss [32] using a VGG16 network [33], mean-square error (MSE), and adversarial loss using a discriminator neural network [34].
