eduzhai > Applied Sciences > Engineering >

Self-Organized Operational Neural Networks for Severe Image Restoration Problems

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Discriminative learning based on convolutional neural networks (CNNs) aims toperform image restoration by learning from training examples of noisy-cleanimage pairs. It has become the go-to methodology for tackling image restorationand has outperformed the traditional non-local class of methods. However, thetop-performing networks are generally composed of many convolutional layers andhundreds of neurons, with trainable parameters in excess of several millions.We claim that this is due to the inherent linear nature of convolution-basedtransformation, which is inadequate for handling severe restoration problems.Recently, a non-linear generalization of CNNs, called the operational neuralnetworks (ONN), has been shown to outperform CNN on AWGN denoising. However,its formulation is burdened by a fixed collection of well-known nonlinearoperators and an exhaustive search to find the best possible configuration fora given architecture, whose efficacy is further limited by a fixed output layeroperator assignment. In this study, we leverage the Taylor series-basedfunction approximation to propose a self-organizing variant of ONNs, Self-ONNs,for image restoration, which synthesizes novel nodal transformations onthe-flyas part of the learning process, thus eliminating the need for redundanttraining runs for operator search. In addition, it enables a finer level ofoperator heterogeneity by diversifying individual connections of the receptivefields and weights. We perform a series of extensive ablation experimentsacross three severe image restoration tasks. Even when a strict equivalence oflearnable parameters is imposed, Self-ONNs surpass CNNs by a considerablemargin across all problems, improving the generalization performance by up to 3dB in terms of PSNR.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...