Adaptive Image Deblurring Convolutional Neural Network with Meta-Tuning

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Motion blur degrades image quality due to sensor-object movement during exposure.
• Deep learning, especially CNNs, shows potential for deblurring but faces limitations like small kernel sizes and dataset overfitting.
• Existing models struggle with generalization to real-world blur domains and often produce artifacts.

Purpose of the Study:

• To develop an advanced deep learning model for effective motion deblurring.
• To enhance the generalization capability of deblurring models across various blur types.
• To address limitations of current CNN-based methods in handling complex motion blur.

Main Methods:

• Proposed the Spatial Feature Selection Network (SFSNet) with a Regional Feature Extractor (RFE) module.
• Introduced the BlurMix dataset with diverse blur types.
• Implemented a meta-tuning strategy for efficient blur domain adaptation.

Main Results:

• SFSNet effectively expands receptive fields and selects critical spatial features for improved deblurring.
• The meta-tuning approach enables rapid adaptation to novel blur distributions with minimal training.
• Experimental results demonstrate significant improvement in deblurring performance and elimination of artifacts across various domains.

Conclusions:

• SFSNet offers a robust solution for motion deblurring, overcoming limitations of traditional CNNs.
• The BlurMix dataset and meta-tuning strategy enhance model generalization and adaptability.
• The proposed method significantly improves deblurring quality and reduces artifacts in real-world applications.