Efficient Image Restoration using Super and Generative Neurons
Adalioglu, Ilke (2024)
Adalioglu, Ilke
2024
Master's Programme in Computing Sciences
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-06-11
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202406077031
https://urn.fi/URN:NBN:fi:tuni-202406077031
Tiivistelmä
Over the last decade, Image restoration methods have utilized Convolutional Neural Networks (CNNs) to map the degraded input to its cleaner approximation. Deep CNNs excel at modeling these complex degradations and achieving high performances with the capacity of numerous hidden layers that can integrate a broader receptive field. This powerful capability comes with the drawback of increased computational and environmental costs, making AI less accessible to research and unsuitable for use in mobile devices and other environments with limited resources. To overcome these limitations, Self Organised Neural Networks (Self-ONNs) have been introduced as a more efficient alternative and better modeling of a biological neuron. Recently proposed Super Neurons empower Self-ONNs (Super-ONNs) by improving their already powerful naturally homogeneous, non-linear operations by increasing their receptive field with non-localised kernels. This key improvement, enables us to use fewer neurons to learn the same or similar performing modeling.
This study aims to develop an efficient Self-ONN model harnessing generative and super neurons, capable of achieving performances comparable to or state-of-the-art performances, with a fraction of parameters. With an extensive set of experiments to study the behavior of Super-ONNs, we demonstrated an understanding of best practices, and using these, we developed more complex models to compete in image restoration problems with a significant decrease in computational complexity.
In shallow model denoising experiments over the SIDD medium dataset, we further improved upon the already high-performing Self-ONN results and obtained the best results using activation-free Super-ONNs with Q = 2, surpassing deep CNNs with a margin of 2 dB and minimal complexity. Competing with deep CNNs, While having 5.6 thousand neurons, our activation-free network achieves 39.67 dB in denoising over the SIDD dataset, with 12 million parameters. Although this model has slightly lower PSNR compared to state-of-the-art methods, it achieves this with significantly fewer parameters, demonstrating high efficiency. In image deblurring, the Self-ONN model falls just short of the best-performing model while having half the number of parameters and FLOPs. Results reveal the potential of our approach to offer highly efficient and competitive models for image restoration tasks, indicating its significance for future advancements.
This study aims to develop an efficient Self-ONN model harnessing generative and super neurons, capable of achieving performances comparable to or state-of-the-art performances, with a fraction of parameters. With an extensive set of experiments to study the behavior of Super-ONNs, we demonstrated an understanding of best practices, and using these, we developed more complex models to compete in image restoration problems with a significant decrease in computational complexity.
In shallow model denoising experiments over the SIDD medium dataset, we further improved upon the already high-performing Self-ONN results and obtained the best results using activation-free Super-ONNs with Q = 2, surpassing deep CNNs with a margin of 2 dB and minimal complexity. Competing with deep CNNs, While having 5.6 thousand neurons, our activation-free network achieves 39.67 dB in denoising over the SIDD dataset, with 12 million parameters. Although this model has slightly lower PSNR compared to state-of-the-art methods, it achieves this with significantly fewer parameters, demonstrating high efficiency. In image deblurring, the Self-ONN model falls just short of the best-performing model while having half the number of parameters and FLOPs. Results reveal the potential of our approach to offer highly efficient and competitive models for image restoration tasks, indicating its significance for future advancements.