Super-Resolution Generator Networks: A comparative study

Modern approaches that tackle super-resolution aim to train a generator network that transforms the low resolution image into a higher resolution one. The core learning capacity of these generator networks is given by stacks of well known image processing blocks such as VGG-16 [SZ14], ResNet[HZRS15] or Inception-v3 [SVI $^{+15]}$ blocks. In the light of recent advancements on the CIFAR-10 [KNH] benchmarks where DenseNet [HLW16] and later SparseNet [ZDD $^{+18]}$ proved superior performance over the architectures that used the formerly mentioned blocks, this paper aims to do a comparative study on the performance changes resulting when using DenseNet or SparseNet blocks in generator networks. We first replicate the results of [JAL16]. This work describes a generator network that uses a stack of four ResNet blocks. This stack is incorporated in two architectures for superresolution, one for x4 magnification and another one for x8. We then proceed and substitute them with DenseNet blocks and SparseNet blocks but keep the same overall training procedure. In order to ensure a fair comparison we adapt the number of blocks for each architecture in order to match the same amount of parameters on all architectures. In all cases the same optimization loss function is used, perceptual loss [JAL16], which for a given image yields a value that is a weighted sum of mean-squared-errors between filters of the target input and generated image evaluated on equivalent convolution layers of the last three blocks in the VGG-16 network (pretrained on the ImageNet [DDS $^{+09]}$ dataset). We monitor on all architectures the loss value, the number of epochs needed to reach the lowest loss, the artifacts generated by each network and the overall appearance of the reconstructions.