Self-Organized Operational Neural Networks for Severe Image Restoration Problems

TL;DR

This paper introduces Self-Organized Operational Neural Networks (Self-ONNs) that dynamically synthesize nonlinear transformations during training, significantly improving severe image restoration performance over traditional CNNs.

Contribution

It proposes a novel self-organizing ONN variant using Taylor series approximation, eliminating the need for operator search and enhancing heterogeneity in image restoration tasks.

Findings

Self-ONNs outperform CNNs by up to 3 dB PSNR in severe restoration tasks.

Self-ONNs require fewer training runs due to on-the-fly operator synthesis.

The method improves generalization across multiple image restoration problems.

Abstract

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