Self-organized operational neural networks for severe image restoration problems

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) dominate image restoration but struggle with severe degradation due to their linear nature.
• Operational Neural Networks (ONNs) offer non-linear alternatives but require extensive operator searches and fixed configurations.
• Existing methods often lack flexibility in handling complex image restoration challenges.

Purpose of the Study:

• To introduce Self-ONNs, a novel self-organizing variant of ONNs for superior image restoration.
• To address the limitations of fixed operator sets and exhaustive searches in traditional ONNs.
• To improve the generalization performance and efficiency of deep learning models in image restoration tasks.

Main Methods:

• Leveraging Taylor series-based function approximation to synthesize novel nodal transformations dynamically.
• Implementing a self-organizing mechanism within ONNs to create transformations on-the-fly during learning.
• Enabling finer operator heterogeneity by diversifying receptive field connections and weights.

Main Results:

• Self-ONNs significantly surpass CNNs in performance across three severe image restoration tasks.
• Performance improvements of up to 3 dB in Peak Signal-to-Noise Ratio (PSNR) were observed.
• The proposed method demonstrates superior generalization capabilities compared to existing CNNs, even with equivalent parameter counts.

Conclusions:

• Self-ONNs represent a significant advancement in image restoration, offering enhanced performance and adaptability.
• The dynamic synthesis of transformations eliminates the need for redundant training and improves model flexibility.
• This approach provides a more effective solution for severe image restoration problems compared to conventional CNNs.