Rediscovery of SSIM index in image reconstruction. SSIM as a loss function.
Default loss function in encoder-decoder based image reconstruction had been L2 loss. Previously, Caffe only provides L2 loss as a built-in loss layer. Generally, L2 loss makes reconstructed image blurry because minimizing L2 loss means maximizing log-likelihood of Gaussian. As you know Gaussian is unimodal.
L1 gains a popularity over L2 because it tends to create less blurry images. However, using either L1 or L2 loss in learning takes enormous time to converge. Both losses are pointwise, error is…

