Outline of FORKLENS shear estimation architecture. The CNN part contains two branches, one fed with PSF and one fed with a galaxy image. The PSF branch has four convolutional layers, each with batch normalization and ReLU activation function. We adopted a 34-layer residual network to extract the information of galaxies where the image is larger and suffers from pixel noise. The two branches are then concatenated following two fully connected layers, where the effect of PSF is corrected. CNN then outputs the galaxy’s properties including size, magnitude, and ellipticities. A further NN calibrates the measured features biased by noise and outputs the final shear estimate. The NN part is in practice a committee of eight independent NNs and the 𝒢₁ (𝒢₂) is the average of the eight outputs.