EHNet: Efficient Hybrid Network with Dual Attention for Image Deblurring.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image blur from motion is a significant challenge in imaging, degrading image quality.
• Deep learning, particularly Convolutional Neural Networks (CNNs) and Transformers, shows promise for image deblurring.
• Existing methods face limitations: CNNs have restricted receptive fields, while Transformers are computationally intensive and lack inductive bias.

Purpose of the Study:

• To develop an efficient deep-learning-based image processing method for effective blur artifact removal.
• To overcome the limitations of existing CNN and Transformer architectures in image deblurring.
• To propose a novel hybrid network that balances local feature extraction and long-range dependency modeling.

Main Methods:

• Proposed an Efficient Hybrid Network (EHNet) utilizing CNN encoders for local features and Transformer decoders for global context.
• Introduced a dual-attention module within the Transformer decoders to capture spatial and channel-wise dependencies.
• Developed a Simple Feature-Embedding Module (SFEM) to reduce computational complexity in the self-attention mechanism.

Main Results:

• The EHNet model demonstrated high-quality image deblurring capabilities.
• Comprehensive experiments on benchmark datasets showed promising quantitative and qualitative results.
• The SFEM significantly reduced computational complexity and memory usage without compromising performance.

Conclusions:

• The proposed EHNet offers an efficient and effective solution for image deblurring.
• The hybrid approach successfully integrates the strengths of CNNs and Transformers.
• This work advances deep learning-based image deblurring with a computationally efficient and high-performing model.