SharpFormer: Learning Local Feature Preserving Global Representations for Image Deblurring.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Conventional Convolutional Neural Networks (CNNs) struggle with non-uniform motion blurs due to limitations in capturing varied local and global image information.
• Existing methods often increase network depth or kernel size, which can be insufficient for severe blurs and fail to address non-uniform blur characteristics.

Purpose of the Study:

• To introduce SharpFormer, a Transformer-based model designed for dynamic scene deblurring.
• To overcome the limitations of CNNs in handling non-uniform blurs and extracting comprehensive local and global features.
• To improve the restoration of details in severely blurred images.

Main Methods:

• Proposed a Transformer-based model, SharpFormer, leveraging long-range dependency learning for motion blur removal.
• Introduced a novel Locality preserving Transformer (LTransformer) block to integrate local information with global features.
• Incorporated a dynamic convolution (DyConv) block to aggregate parallel kernels for handling non-uniform blurs.
• Utilized a two-stage attentive framework combining global, hybrid, and local feature extraction.

Main Results:

• SharpFormer demonstrated superior performance in blurred image restoration compared to state-of-the-art methods.
• The model effectively handled large blur variations and non-uniform blur characteristics.
• Experiments on GoPro and REDS datasets validated the effectiveness of the proposed approach.

Conclusions:

• The proposed SharpFormer model offers a significant advancement in dynamic scene deblurring.
• Transformer-based architectures, with tailored blocks like LTransformer and DyConv, are highly effective for complex image restoration tasks.
• SharpFormer provides a robust solution for restoring details in severely blurred images, outperforming existing techniques.