Deep Learning and Neural Architecture Search for Optimizing Binary Neural Network Image Super Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning significantly advanced super-resolution (SR) technology.
• Resource-constrained devices require efficient SR models with low computational and memory footprints.
• Binary Neural Networks (BNNs) offer reduced complexity but need effective architectures.

Purpose of the Study:

• To develop an efficient neural architecture search (NAS) method for Binary Neural Networks (BNNs) specifically for image super-resolution (SR).
• To address the challenge of designing effective BNN architectures for SR tasks efficiently.

Main Methods:

• A novel differentiable NAS approach tailored for the SR task's requirements.
• Adaptation of the search space to optimize for SR performance.
• Integration of Libra Parameter Binarization (Libra-PB) to enhance information retention.

Main Results:

• Generated network architectures require only one-third of the parameters compared to conventional methods.
• Achieved comparable performance to existing SR models despite significant parameter reduction.
• Demonstrated the efficiency and effectiveness of the proposed binary network search method.

Conclusions:

• The proposed method efficiently designs compact and performant BNNs for image super-resolution.
• This approach facilitates the deployment of advanced SR capabilities on resource-limited devices.
• Libra-PB integration aids in preserving crucial information during binarization for SR.