Image deblurring method and system based on frequency domain self-attention mechanism
By using an asymmetric network architecture with a frequency domain self-attention mechanism, the limitations of existing image deblurring methods are overcome, achieving efficient and accurate image deblurring results that can adapt to image restoration of different blur types.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING XIAOYING TECH CO LTD
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-17
AI Technical Summary
Existing image deblurring methods suffer from problems such as local invariance of convolution operations, need for prior knowledge of the blur type, and information loss due to reduced resolution when dealing with motion blur and out-of-focus blur, which affect the quality and efficiency of image restoration.
A frequency-domain-based self-attention mechanism is adopted. Through the asymmetric architecture of self-attention solver and discriminative feedforward network, it is embedded in the decoder and encoder modules. Frequency domain operations are used to estimate the scaling dot product attention and frequency information, forming an end-to-end trainable network that adaptively determines the frequency information.
It eliminates the need for extensive sample collection, reducing computational complexity and time costs, and improving the accuracy and efficiency of image deblurring, effectively restoring clear images.
Smart Images
Figure CN122415385A_ABST